[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-anthropic-math-olympiad":3,"mdc--41qe8-key":34,"related-repo-anthropic-math-olympiad":1420,"related-org-anthropic-math-olympiad":1528},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":12,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":29,"sourceUrl":32,"mdContent":33},"math-olympiad","solve olympiad math problems","Solve competition math problems (IMO, Putnam, USAMO, AIME) with adversarial verification that catches the errors self-verification misses. Activates when asked to 'solve this IMO problem', 'prove this olympiad inequality', 'verify this competition proof', 'find a counterexample', 'is this proof correct', or for any problem with 'IMO', 'Putnam', 'USAMO', 'olympiad', or 'competition math' in it. Uses pure reasoning (no tools) — then a fresh-context adversarial verifier attacks the proof using specific failure patterns, not generic 'check logic'. Outputs calibrated confidence — will say 'no confident solution' rather than bluff. If LaTeX is available, produces a clean PDF after verification passes.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},"anthropic","Anthropic","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fanthropic.png","anthropics",[13,17],{"name":14,"slug":15,"type":16},"Research","research","tag",{"name":18,"slug":19,"type":16},"Mathematics","mathematics",32228,"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fclaude-plugins-official","2026-04-06T18:00:38.100685",null,3591,[26,27,28],"claude-code","mcp","skills",{"repoUrl":21,"stars":20,"forks":24,"topics":30,"description":31},[26,27,28],"Official, Anthropic-managed directory of high quality Claude Code Plugins.","https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fclaude-plugins-official\u002Ftree\u002FHEAD\u002Fplugins\u002Fmath-olympiad\u002Fskills\u002Fmath-olympiad","---\nname: math-olympiad\ndescription:\n  \"Solve competition math problems (IMO, Putnam, USAMO, AIME) with adversarial\n  verification that catches the errors self-verification misses. Activates when\n  asked to 'solve this IMO problem', 'prove this olympiad inequality', 'verify\n  this competition proof', 'find a counterexample', 'is this proof correct', or\n  for any problem with 'IMO', 'Putnam', 'USAMO', 'olympiad', or 'competition\n  math' in it. Uses pure reasoning (no tools) — then a fresh-context adversarial\n  verifier attacks the proof using specific failure patterns, not generic 'check\n  logic'. Outputs calibrated confidence — will say 'no confident solution'\n  rather than bluff. If LaTeX is available, produces a clean PDF after\n  verification passes.\"\nversion: 0.1.0\n---\n\n# Math Olympiad Solver\n\n## The five things that change outcomes\n\n1. **Strip thinking before verifying** — a verifier that sees the reasoning is\n   biased toward agreement. Fresh context, cleaned proof only.\n2. **\"Does this prove RH?\"** — if your theorem's specialization to ζ is a famous\n   open problem, you have a gap. Most reliable red flag.\n3. **Short proof → extract the general lemma** — try 2×2 counterexamples. If\n   general form is false, find what's special about THIS instance.\n4. **Same gap twice → step back** — the case split may be obscuring a unified\n   argument. Three lines sometimes does what twelve pages couldn't.\n5. **Say \"no confident solution\"** — wrong-and-confident is worse than honest\n   abstain.\n\n---\n\n**Tool policy**: Solvers and verifiers use THINKING ONLY in the tight-budget\nworkflow. Competition math is reasoning. Computation is for deep mode (§6c), and\neven then bounded — a recurrence that's doubly-exponential can't be computed\npast n~30, work mod 2^m instead.\n\n---\n\n## When to use which approach\n\n| Problem                                              | Approach                                                                       | Verification              |\n| ---------------------------------------------------- | ------------------------------------------------------------------------------ | ------------------------- |\n| AIME numeric answer                                  | Best-of-N → majority vote                                                      | Answer check only         |\n| Olympiad proof (IMO\u002FPutnam\u002FUSAMO)                    | Full workflow below                                                            | 5-pass adversarial        |\n| \"Is this proof correct?\"                             | Skip to verification (step 4)                                                  | Adversarial + spec-gaming |\n| **Full problem set** (e.g. all 6 from a competition) | Sequential: one full workflow per problem, collect results, compile single PDF | Per-problem adversarial   |\n\n**Batch in one Workflow**: Set `opts.label` on every `agent()` call to include\nthe problem ID (e.g., `label: \"P3:solver:2\"`). Without labels, 36 results come\nback with no problem association. Run problems in parallel — the label is what\nmatters, not ordering.\n\n### For a full problem set\n\nLaunch one solver workflow per problem (same VERBATIM prompt, different\nstatement). Run them in parallel. When all return, run adversarial verification\nper problem. Problems that pass get their proof in the PDF; problems that\nabstain get \"No confident solution\" with partial notes.\n\nDon't try to solve all N problems in one agent's context — each problem needs\nits own thinking budget and its own fresh-context verifier. The composition is\nmechanical: collect the per-problem outputs, fill in LaTeX sections, compile\nonce. | \"Simplify this proof\" | Skip to presentation (step 8) | — |\n\n---\n\n## The Workflow\n\n### 1. Interpretation check (30 seconds, catches 50\u002F63 of one class of errors)\n\nBefore solving anything, identify the interpretation.\n\n> Read the problem statement. List 2-3 ways it could be interpreted. For each:\n> is this reading TRIVIAL? If one reading makes the problem easy and another\n> makes it hard, the hard one is almost certainly intended. State which\n> interpretation you're solving and WHY you believe it's the intended one.\n\nThe Aletheia case study found 50 of 63 \"technically correct\" solutions were for\nthe wrong interpretation. Olympiad problems often have a trap easy reading.\n\n### 2. Generate candidates with internal refinement (parallel, thinking only)\n\nLaunch 8-12 attempt agents in parallel. **Each agent internally iterates** —\nsolve → self-improve → self-verify → correct → repeat. This is the Yang-Huang\nstructure that achieves 85.7% on IMO: one-shot solving isn't enough; per-attempt\nrefinement matters.\n\n**The Agent tool cannot enforce tool restriction.** Subagents get the full tool\nset. The only mechanism is the prompt. Use this prompt VERBATIM — do not\nsummarize, do not synthesize your own:\n\n```\nNO COMPUTATION. Do not use Bash, Python, WebSearch, Read, Write, or any tool that runs code or fetches data. Numerical verification is not a proof step. \"I computed n=1..10 and the pattern holds\" is not a proof.\n\n(If your agent harness requires a StructuredOutput or similar return-mechanism tool call, that is NOT a computation tool — call it to return your answer. The restriction is on tools that DO work, not tools that REPORT work.)\n\nYour internal process (iterate until done):\n- Solve: Complete rigorous solution.\n- Self-improve: Reread. Fix gaps before a grader sees it.\n- Self-verify: Strict grader mode. Every step justified?\n- Correct: Fix and re-verify. Up to 5 rounds.\n- Stop: Self-verify passes twice clean, OR 5 rounds, OR approach fundamentally wrong.\n\nA correct answer from flawed reasoning is a failure. If incomplete, say so honestly. Never hide gaps.\n\nPROBLEM: \u003Cinsert the problem statement here>\nANGLE: \u003Cinsert one starting angle here>\n```\n\nThe first two paragraphs are load-bearing. A session that writes its own prompt\nand omits them will produce subagents that grind Python for 30 iterations and\nconfidently get wrong answers — a pattern that fits n≤10 but fails at n=100 is\nnot a proof.\n\nStarting angles (vary across agents — see `references\u002Fsolver_heuristics.md`):\n\n- Work out small cases (test past n=3)\n- Look for an invariant or monovariant\n- Consider the extremal case\n- Try induction\n- What symmetries?\n- Work backwards\n- Drop a condition — where does it become trivially false?\n- Generalize (inventor's paradox — more structure is sometimes easier)\n\nEach returns its FINAL state (not intermediate rounds):\n\n```\n**Verdict**: complete solution | partial result | no progress\n**Rounds**: [how many verify→correct cycles]\n**Method**: [key idea, one paragraph]\n**Detailed Solution**: [full step-by-step, every step justified]\n**Answer**: [if applicable]\n**Self-verification notes**: [what you caught and fixed; remaining concerns]\n```\n\n**Retry policy**: If an agent fails or times out, retry once. Transient failures\nhappen.\n\n### 3. Clean the solution (context isolation — the #1 lever)\n\nThe thinking trace biases the verifier toward agreement — a long chain of\nreasoning reads as supporting evidence even when the conclusion is wrong. Before\nany verification, strip:\n\n- All thinking-block content\n- All \"Let me try...\" \u002F \"Actually wait...\" \u002F \"Hmm\" prose\n- All false starts and backtracking\n\nWhat remains: problem statement + clean final argument only.\n\nExtract only the **Method** + **Proof** + **Answer** sections from each solver's\noutput. The verifier never sees how the solver got there.\n\n### 4. Adversarial verify (fresh context, pattern-armed)\n\nFor each cleaned solution, launch a fresh verifier agent. **Fresh context**: it\nsees only (problem statement + cleaned solution). **No tools.**\n\nThe verifier's job is to ATTACK, not grade. Load\n`references\u002Fadversarial_prompts.md` for the prompts. The key patterns it runs:\n\n| Pattern | The check                                                                                                                                                          |\n| ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------ |\n| **#4**  | Does this theorem specialize to a famous object (ζ, quadratic reciprocity, etc.) and prove something open about it? → gap                                          |\n| **#18** | Substitute the proof's own intermediate identities into any \"remaining gap.\" Recover the original claim? → tautological                                            |\n| **#40** | Is any step a \"one-line lemma\"? Extract the GENERAL form. Find a 2×2 counterexample. If the general form is false, find what special structure saves THIS instance |\n| **#5**  | For each invoked theorem: re-check hypotheses FROM SCRATCH. \"Continuous on [0,1]\" ≠ \"continuous on ℝ\"                                                              |\n| **#6**  | Any infinite sum \"bounded\" via a regularized value? Check the boundary — if there's a pole there, the sum diverges                                                 |\n\nFull pattern list: `references\u002Fverifier_patterns.md`\n\nVerifier returns:\n\n```\n**Verdict**: HOLDS | HOLE FOUND | UNCLEAR\n\n**If HOLE FOUND**:\n- Location: [quote the problematic step]\n- Pattern: [which check fired, or \"other\"]\n- Why it breaks: [specific]\n- Fixable?: [yes with X \u002F no, fundamental]\n```\n\n### 5. Rank and vote-verify (asymmetric + early exit)\n\nRank solutions by (verdict, verifier confidence). Take the top one. Run up to 5\nfresh verifier agents.\n\n**Asymmetric thresholds**: 4 HOLDS to confirm, 2 HOLE FOUND to refute. Why\nasymmetric: one flaky verifier shouldn't kill a correct proof; but two\nindependent dissents is a real signal.\n\n**Pigeonhole early exit**: stop launching verifiers once the outcome is decided.\n\n- 2 say HOLE FOUND → refuted, stop (save the remaining 3 calls)\n- 4 say HOLDS → confirmed, stop (save the 5th)\n- After 3 verifiers: if 2 HOLDS + 1 HOLE, launch 2 more (outcome undecided). If\n  3 HOLDS + 0 HOLE, launch 1 more (could still hit 4-1).\n\n**Dual context-isolation**: each verifier is blind to (a) the solver's thinking\ntrace — already stripped in step 3 — AND (b) other verifiers' verdicts. Each\nverifier thinks it's the first. No \"3 agents already confirmed this\" social\nproof.\n\n**A solver cannot verify its own solution.** Different agent, fresh context.\n\n### 5b. When one case won't close — step back before grinding\n\nIf a proof splits into cases and one case proves easily but the other resists:\n**before grinding through the hard case, ask whether there's a route that makes\nthe split disappear.**\n\nThe pattern that saves you: the hard case's very hypothesis often implies\nsomething strong about an _intermediate object_ you haven't looked at. Use that\nimplication directly instead of the original chain.\n\nConcrete shape: proving f(n) ≤ cn for a constrained function f, with a case\nsplit on a prime p dividing f(n). One branch closes by index arguments in\n(ℤ\u002Fp^e)\\*. The other branch resists — same group structure, but the arithmetic\ndoesn't contradict. The fix: the hypothesis \"p | f(n)\" plugged back into the\ngoverning equation implies **f(p) = p itself**. Once you have that, a\nFermat+Dirichlet argument kills both branches in three lines. The case split was\na detour — it was splitting on a variable that, under the hypothesis, takes a\nknown value.\n\nCheck when stuck on case B:\n\n- What does case B's hypothesis imply about f at _other_ inputs?\n- Is there a different pair (a,b) to plug into the governing equation?\n- Are you proving too much? (A cleaner contradiction needs less machinery.)\n\nThis is also a presentation-pass win: the split-free proof is shorter AND more\ngeneral.\n\n### 6. Revise (if needed)\n\nIf verification finds a hole: launch a reviser agent. It gets (cleaned\nsolution + verifier's hole report). STILL no access to the original thinking —\nthe reviser works from the hole, not by rereading how you got there.\n\n```\nA verifier found this issue in the proof:\n[hole report]\n\nFix the proof. If the hole is fundamental (the approach doesn't work), say so and return **Verdict: no confident solution** with what partial progress remains.\n\nFor any step you cannot fully close, mark it inline: [GAP: specific description of what remains]. Gaps in the proof text, not in a separate list — they're greppable and the next reviser knows exactly where to look.\n```\n\nUp to 3 revise cycles. Then re-run the vote on the revised proof.\n\n**If pattern #40 fired** (one-line-proof-too-clean), the reviser gets a stronger\nbrief — the Adversarial Brief template from `references\u002Fadversarial_prompts.md`\n§7. It forces a binary: \"the general lemma is obviously false (here's a 2×2\ncounterexample) — so either find what's special about THIS case, or find where\nthe proof breaks.\" Can't return \"looks fine.\"\n\n### 6c. Deep mode (when tight-budget abstains)\n\nThe standard workflow is tight-budget: 8 solvers, ~15 min, pure reasoning. When\nit abstains, the problem may need more time, not more capability.\n\n**Deep mode** is a single focused agent with:\n\n- **Unlimited time** — no wall-clock pressure\n- **Targeted computation allowed** — modular arithmetic checks, small-case\n  enumeration, symbolic verification of identities. NOT exploratory brute force\n  or unbounded recursion.\n- **The abstention reason as starting point** — if verifiers found a specific\n  gap, start there. If solvers never claimed complete, start from what they\n  partially proved.\n\nThe archetype: a focused agent that gets the proven-so-far state plus \"one case\nof Lemma 5 is open\" — and finds a 3-line argument the case split was obscuring.\nOften under 10 minutes with almost no computation. Deep mode is about giving the\nproblem sustained attention, not throwing compute at it.\n\n**What deep mode is NOT**: open-ended exploration, literature search, looking up\nsolutions, multi-day investigation. That's a different workflow\n(`math-research`). Deep mode is still \"solve THIS problem yourself\" — just\nwithout the clock.\n\n**NO WEB. NO LOOKUP.** Deep mode may use Bash\u002FPython for bounded computation,\nbut NEVER WebFetch, WebSearch, or any network access. Finding the solution on\nAoPS or a blog is not solving the problem — it's cheating on an olympiad, and it\nteaches us nothing about the skill's actual capability. Put this at the TOP of\nthe deep-mode prompt:\n\n```\nNO WEB ACCESS. Do not use WebFetch, WebSearch, or any tool that touches the internet. Do not look up this problem, its solution, or related problems. You are solving this yourself — the only allowed computation is local (Bash\u002FPython for mod-k arithmetic, small-case enumeration n≤10, symbolic identity checks). If you invoke a web tool, the proof is void.\n```\n\n**Computation bounds in deep mode** (bug #8 lesson): A6's b\\_{n+1}=2b_n²+b_n+1\nis doubly-exponential; b_99 has ~10^{2^98} digits. Never compute such objects\nexactly — work in ℤ\u002F2^m, or track only v_p(·), or prove the recursion mod the\nquantity you care about. If a computation is running longer than 60 seconds,\nit's probably unbounded. Kill it and work symbolically.\n\n**Step 6d (not optional)**: After any ABSTAIN at the verify stage, automatically\nlaunch one deep-mode agent before writing the abstention into the output. Give\nit:\n\n- The problem statement\n- The best partial proof from tight-budget solvers\n- The verifier gap descriptions (what specifically didn't close)\n- The instruction: \"NO WEB ACCESS — do not look up this problem or its solution.\n  Bounded local computation allowed (mod 2^k, small cases n≤10, symbolic\n  identity checks via Bash\u002FPython only). 60-second computation limit. If n≤10\n  brute force reveals a pattern the tight-budget solvers missed, that pattern IS\n  the proof structure.\"\n\nThe deep agent may find the construction the pure-reasoning solvers couldn't\nsee. If it also abstains, THEN write the abstention. Do not skip this step —\nproblems with √n or log n answers are often invisible to pure reasoning because\nthe optimal structure is the asymmetric one.\n\n**Orchestrator self-restraint**: The orchestrator itself must not web-search the\nproblem \"to help\" the deep agent. If you're tempted to Fetch an AoPS thread\n\"just to check the answer,\" don't — that contaminates the skill's output and\nmisrepresents its capability.\n\n### 7. Calibrated abstention\n\nIf 3 revise cycles all fail: **stop and admit it.**\n\n```\n**Verdict**: no confident solution\n\n**What was tried**: [approaches]\n**What WAS proven**: [any lemma or partial result that survived verification]\n**Where it breaks**: [the unfixed hole]\n```\n\nDo NOT guess. A wrong confident answer is worse than an honest \"couldn't solve\nit.\" The metric that matters is CONDITIONAL accuracy — when you say \"solved,\"\nare you right?\n\n### 8. Presentation pass (after correctness is established)\n\nA VERIFIED-CORRECT proof is often not a BEAUTIFUL proof. The order you\ndiscovered it is rarely the best order to present it. Launch a fresh\npresentation agent with the verified proof.\n\nLoad `references\u002Fpresentation_prompts.md`. The agent asks:\n\n- What's the simplest way to say this?\n- Which lemmas should be inlined? Which deserve to stand alone?\n- Is anything OVERKILL? (constructing a double exponential when linear suffices)\n- Now that we know the answer, is there a 3-line hindsight proof?\n\nOutput: LaTeX-formatted proof. If `pdflatex` is available\n(`scripts\u002Fcheck_latex.sh` returns 0), also compile to PDF via\n`scripts\u002Fcompile_pdf.sh`.\n\n---\n\n## Model tier defaults\n\nRead `references\u002Fmodel_tier_defaults.md` for full details. Summary:\n\n| Model  | Solvers | Verify passes          | Abstain after  | Presentation           |\n| ------ | ------- | ---------------------- | -------------- | ---------------------- |\n| Haiku  | 8       | 3                      | 2 revise fails | skip                   |\n| Sonnet | 4       | 5                      | 3 revise fails | yes                    |\n| Opus   | 3       | 5 + full pattern sweep | 4 revise fails | 2 drafts, pick cleaner |\n\nWeaker models: more parallel attempts, faster abstention. Stronger models:\ndeeper verification, more presentation effort.\n\n---\n\n## For numeric-answer problems (AIME-style)\n\nSkip the proof machinery. Run 5-7 solvers with varied approaches, take majority\nvote on the numeric answer. If no majority: verify the top 2 candidates by\nsubstitution.\n\n---\n\n## Key references\n\n- `references\u002Fverifier_patterns.md` — the 12 adversarial checks\n- `references\u002Fadversarial_prompts.md` — ready-to-use verifier prompts\n- `references\u002Fpresentation_prompts.md` — beautification prompts + LaTeX template\n- `references\u002Fmodel_tier_defaults.md` — per-model configuration\n\n---\n\n## What makes this different from generic verify-and-refine\n\n1. **Dual context isolation**: verifier is blind to (a) the solver's thinking\n   trace — which biases toward agreement — and (b) other verifiers' verdicts —\n   social proof also biases. Each verifier thinks it's first.\n2. **Pattern-specific attacks**: not \"is this correct?\" but \"does this make the\n   #40 mistake? the #4 mistake?\" Specific beats generic. The 7-category\n   refutation taxonomy gives the verifier a checklist.\n3. **Asymmetric vote + pigeonhole exit**: 4-to-confirm, 2-to-refute. One flaky\n   verifier doesn't kill a correct proof; two dissents does. Stop launching\n   verifiers once the outcome is decided — saves ~30% of verification cost on\n   clear cases.\n4. **Specification-gaming check first**: explicitly asks \"is this the intended\n   interpretation?\" before solving. The #1 failure mode in prior work (50\u002F63\n   \"correct\" answers solved the wrong reading).\n5. **Calibrated abstention**: will say \"no confident solution\" with partial\n   results. Optimizes conditional accuracy, not coverage.\n6. **Presentation pass**: correctness and elegance are separate steps. The\n   presentation agent gets the VERIFIED proof and finds the cleanest way to say\n   it.\n",{"data":35,"body":37},{"name":4,"description":6,"version":36},"0.1.0",{"type":38,"children":39},"root",[40,49,56,112,116,127,130,136,246,281,288,293,298,301,307,313,318,327,332,338,350,360,372,377,390,434,439,448,458,464,469,487,492,517,523,540,553,663,674,679,688,694,699,709,719,737,747,757,763,773,786,798,803,828,833,839,844,853,858,875,881,886,896,929,934,952,962,971,981,991,1014,1019,1029,1035,1045,1054,1059,1065,1070,1083,1106,1135,1138,1144,1157,1277,1282,1285,1291,1296,1299,1305,1348,1351,1357],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"math-olympiad-solver",[46],{"type":47,"value":48},"text","Math Olympiad Solver",{"type":41,"tag":50,"props":51,"children":53},"h2",{"id":52},"the-five-things-that-change-outcomes",[54],{"type":47,"value":55},"The five things that change outcomes",{"type":41,"tag":57,"props":58,"children":59},"ol",{},[60,72,82,92,102],{"type":41,"tag":61,"props":62,"children":63},"li",{},[64,70],{"type":41,"tag":65,"props":66,"children":67},"strong",{},[68],{"type":47,"value":69},"Strip thinking before verifying",{"type":47,"value":71}," — a verifier that sees the reasoning is\nbiased toward agreement. Fresh context, cleaned proof only.",{"type":41,"tag":61,"props":73,"children":74},{},[75,80],{"type":41,"tag":65,"props":76,"children":77},{},[78],{"type":47,"value":79},"\"Does this prove RH?\"",{"type":47,"value":81}," — if your theorem's specialization to ζ is a famous\nopen problem, you have a gap. Most reliable red flag.",{"type":41,"tag":61,"props":83,"children":84},{},[85,90],{"type":41,"tag":65,"props":86,"children":87},{},[88],{"type":47,"value":89},"Short proof → extract the general lemma",{"type":47,"value":91}," — try 2×2 counterexamples. If\ngeneral form is false, find what's special about THIS instance.",{"type":41,"tag":61,"props":93,"children":94},{},[95,100],{"type":41,"tag":65,"props":96,"children":97},{},[98],{"type":47,"value":99},"Same gap twice → step back",{"type":47,"value":101}," — the case split may be obscuring a unified\nargument. Three lines sometimes does what twelve pages couldn't.",{"type":41,"tag":61,"props":103,"children":104},{},[105,110],{"type":41,"tag":65,"props":106,"children":107},{},[108],{"type":47,"value":109},"Say \"no confident solution\"",{"type":47,"value":111}," — wrong-and-confident is worse than honest\nabstain.",{"type":41,"tag":113,"props":114,"children":115},"hr",{},[],{"type":41,"tag":117,"props":118,"children":119},"p",{},[120,125],{"type":41,"tag":65,"props":121,"children":122},{},[123],{"type":47,"value":124},"Tool policy",{"type":47,"value":126},": Solvers and verifiers use THINKING ONLY in the tight-budget\nworkflow. Competition math is reasoning. Computation is for deep mode (§6c), and\neven then bounded — a recurrence that's doubly-exponential can't be computed\npast n~30, work mod 2^m instead.",{"type":41,"tag":113,"props":128,"children":129},{},[],{"type":41,"tag":50,"props":131,"children":133},{"id":132},"when-to-use-which-approach",[134],{"type":47,"value":135},"When to use which approach",{"type":41,"tag":137,"props":138,"children":139},"table",{},[140,164],{"type":41,"tag":141,"props":142,"children":143},"thead",{},[144],{"type":41,"tag":145,"props":146,"children":147},"tr",{},[148,154,159],{"type":41,"tag":149,"props":150,"children":151},"th",{},[152],{"type":47,"value":153},"Problem",{"type":41,"tag":149,"props":155,"children":156},{},[157],{"type":47,"value":158},"Approach",{"type":41,"tag":149,"props":160,"children":161},{},[162],{"type":47,"value":163},"Verification",{"type":41,"tag":165,"props":166,"children":167},"tbody",{},[168,187,205,223],{"type":41,"tag":145,"props":169,"children":170},{},[171,177,182],{"type":41,"tag":172,"props":173,"children":174},"td",{},[175],{"type":47,"value":176},"AIME numeric answer",{"type":41,"tag":172,"props":178,"children":179},{},[180],{"type":47,"value":181},"Best-of-N → majority vote",{"type":41,"tag":172,"props":183,"children":184},{},[185],{"type":47,"value":186},"Answer check only",{"type":41,"tag":145,"props":188,"children":189},{},[190,195,200],{"type":41,"tag":172,"props":191,"children":192},{},[193],{"type":47,"value":194},"Olympiad proof (IMO\u002FPutnam\u002FUSAMO)",{"type":41,"tag":172,"props":196,"children":197},{},[198],{"type":47,"value":199},"Full workflow below",{"type":41,"tag":172,"props":201,"children":202},{},[203],{"type":47,"value":204},"5-pass adversarial",{"type":41,"tag":145,"props":206,"children":207},{},[208,213,218],{"type":41,"tag":172,"props":209,"children":210},{},[211],{"type":47,"value":212},"\"Is this proof correct?\"",{"type":41,"tag":172,"props":214,"children":215},{},[216],{"type":47,"value":217},"Skip to verification (step 4)",{"type":41,"tag":172,"props":219,"children":220},{},[221],{"type":47,"value":222},"Adversarial + spec-gaming",{"type":41,"tag":145,"props":224,"children":225},{},[226,236,241],{"type":41,"tag":172,"props":227,"children":228},{},[229,234],{"type":41,"tag":65,"props":230,"children":231},{},[232],{"type":47,"value":233},"Full problem set",{"type":47,"value":235}," (e.g. all 6 from a competition)",{"type":41,"tag":172,"props":237,"children":238},{},[239],{"type":47,"value":240},"Sequential: one full workflow per problem, collect results, compile single PDF",{"type":41,"tag":172,"props":242,"children":243},{},[244],{"type":47,"value":245},"Per-problem adversarial",{"type":41,"tag":117,"props":247,"children":248},{},[249,254,256,263,265,271,273,279],{"type":41,"tag":65,"props":250,"children":251},{},[252],{"type":47,"value":253},"Batch in one Workflow",{"type":47,"value":255},": Set ",{"type":41,"tag":257,"props":258,"children":260},"code",{"className":259},[],[261],{"type":47,"value":262},"opts.label",{"type":47,"value":264}," on every ",{"type":41,"tag":257,"props":266,"children":268},{"className":267},[],[269],{"type":47,"value":270},"agent()",{"type":47,"value":272}," call to include\nthe problem ID (e.g., ",{"type":41,"tag":257,"props":274,"children":276},{"className":275},[],[277],{"type":47,"value":278},"label: \"P3:solver:2\"",{"type":47,"value":280},"). Without labels, 36 results come\nback with no problem association. Run problems in parallel — the label is what\nmatters, not ordering.",{"type":41,"tag":282,"props":283,"children":285},"h3",{"id":284},"for-a-full-problem-set",[286],{"type":47,"value":287},"For a full problem set",{"type":41,"tag":117,"props":289,"children":290},{},[291],{"type":47,"value":292},"Launch one solver workflow per problem (same VERBATIM prompt, different\nstatement). Run them in parallel. When all return, run adversarial verification\nper problem. Problems that pass get their proof in the PDF; problems that\nabstain get \"No confident solution\" with partial notes.",{"type":41,"tag":117,"props":294,"children":295},{},[296],{"type":47,"value":297},"Don't try to solve all N problems in one agent's context — each problem needs\nits own thinking budget and its own fresh-context verifier. The composition is\nmechanical: collect the per-problem outputs, fill in LaTeX sections, compile\nonce. | \"Simplify this proof\" | Skip to presentation (step 8) | — |",{"type":41,"tag":113,"props":299,"children":300},{},[],{"type":41,"tag":50,"props":302,"children":304},{"id":303},"the-workflow",[305],{"type":47,"value":306},"The Workflow",{"type":41,"tag":282,"props":308,"children":310},{"id":309},"_1-interpretation-check-30-seconds-catches-5063-of-one-class-of-errors",[311],{"type":47,"value":312},"1. Interpretation check (30 seconds, catches 50\u002F63 of one class of errors)",{"type":41,"tag":117,"props":314,"children":315},{},[316],{"type":47,"value":317},"Before solving anything, identify the interpretation.",{"type":41,"tag":319,"props":320,"children":321},"blockquote",{},[322],{"type":41,"tag":117,"props":323,"children":324},{},[325],{"type":47,"value":326},"Read the problem statement. List 2-3 ways it could be interpreted. For each:\nis this reading TRIVIAL? If one reading makes the problem easy and another\nmakes it hard, the hard one is almost certainly intended. State which\ninterpretation you're solving and WHY you believe it's the intended one.",{"type":41,"tag":117,"props":328,"children":329},{},[330],{"type":47,"value":331},"The Aletheia case study found 50 of 63 \"technically correct\" solutions were for\nthe wrong interpretation. Olympiad problems often have a trap easy reading.",{"type":41,"tag":282,"props":333,"children":335},{"id":334},"_2-generate-candidates-with-internal-refinement-parallel-thinking-only",[336],{"type":47,"value":337},"2. Generate candidates with internal refinement (parallel, thinking only)",{"type":41,"tag":117,"props":339,"children":340},{},[341,343,348],{"type":47,"value":342},"Launch 8-12 attempt agents in parallel. ",{"type":41,"tag":65,"props":344,"children":345},{},[346],{"type":47,"value":347},"Each agent internally iterates",{"type":47,"value":349}," —\nsolve → self-improve → self-verify → correct → repeat. This is the Yang-Huang\nstructure that achieves 85.7% on IMO: one-shot solving isn't enough; per-attempt\nrefinement matters.",{"type":41,"tag":117,"props":351,"children":352},{},[353,358],{"type":41,"tag":65,"props":354,"children":355},{},[356],{"type":47,"value":357},"The Agent tool cannot enforce tool restriction.",{"type":47,"value":359}," Subagents get the full tool\nset. The only mechanism is the prompt. Use this prompt VERBATIM — do not\nsummarize, do not synthesize your own:",{"type":41,"tag":361,"props":362,"children":366},"pre",{"className":363,"code":365,"language":47},[364],"language-text","NO COMPUTATION. Do not use Bash, Python, WebSearch, Read, Write, or any tool that runs code or fetches data. Numerical verification is not a proof step. \"I computed n=1..10 and the pattern holds\" is not a proof.\n\n(If your agent harness requires a StructuredOutput or similar return-mechanism tool call, that is NOT a computation tool — call it to return your answer. The restriction is on tools that DO work, not tools that REPORT work.)\n\nYour internal process (iterate until done):\n- Solve: Complete rigorous solution.\n- Self-improve: Reread. Fix gaps before a grader sees it.\n- Self-verify: Strict grader mode. Every step justified?\n- Correct: Fix and re-verify. Up to 5 rounds.\n- Stop: Self-verify passes twice clean, OR 5 rounds, OR approach fundamentally wrong.\n\nA correct answer from flawed reasoning is a failure. If incomplete, say so honestly. Never hide gaps.\n\nPROBLEM: \u003Cinsert the problem statement here>\nANGLE: \u003Cinsert one starting angle here>\n",[367],{"type":41,"tag":257,"props":368,"children":370},{"__ignoreMap":369},"",[371],{"type":47,"value":365},{"type":41,"tag":117,"props":373,"children":374},{},[375],{"type":47,"value":376},"The first two paragraphs are load-bearing. A session that writes its own prompt\nand omits them will produce subagents that grind Python for 30 iterations and\nconfidently get wrong answers — a pattern that fits n≤10 but fails at n=100 is\nnot a proof.",{"type":41,"tag":117,"props":378,"children":379},{},[380,382,388],{"type":47,"value":381},"Starting angles (vary across agents — see ",{"type":41,"tag":257,"props":383,"children":385},{"className":384},[],[386],{"type":47,"value":387},"references\u002Fsolver_heuristics.md",{"type":47,"value":389},"):",{"type":41,"tag":391,"props":392,"children":393},"ul",{},[394,399,404,409,414,419,424,429],{"type":41,"tag":61,"props":395,"children":396},{},[397],{"type":47,"value":398},"Work out small cases (test past n=3)",{"type":41,"tag":61,"props":400,"children":401},{},[402],{"type":47,"value":403},"Look for an invariant or monovariant",{"type":41,"tag":61,"props":405,"children":406},{},[407],{"type":47,"value":408},"Consider the extremal case",{"type":41,"tag":61,"props":410,"children":411},{},[412],{"type":47,"value":413},"Try induction",{"type":41,"tag":61,"props":415,"children":416},{},[417],{"type":47,"value":418},"What symmetries?",{"type":41,"tag":61,"props":420,"children":421},{},[422],{"type":47,"value":423},"Work backwards",{"type":41,"tag":61,"props":425,"children":426},{},[427],{"type":47,"value":428},"Drop a condition — where does it become trivially false?",{"type":41,"tag":61,"props":430,"children":431},{},[432],{"type":47,"value":433},"Generalize (inventor's paradox — more structure is sometimes easier)",{"type":41,"tag":117,"props":435,"children":436},{},[437],{"type":47,"value":438},"Each returns its FINAL state (not intermediate rounds):",{"type":41,"tag":361,"props":440,"children":443},{"className":441,"code":442,"language":47},[364],"**Verdict**: complete solution | partial result | no progress\n**Rounds**: [how many verify→correct cycles]\n**Method**: [key idea, one paragraph]\n**Detailed Solution**: [full step-by-step, every step justified]\n**Answer**: [if applicable]\n**Self-verification notes**: [what you caught and fixed; remaining concerns]\n",[444],{"type":41,"tag":257,"props":445,"children":446},{"__ignoreMap":369},[447],{"type":47,"value":442},{"type":41,"tag":117,"props":449,"children":450},{},[451,456],{"type":41,"tag":65,"props":452,"children":453},{},[454],{"type":47,"value":455},"Retry policy",{"type":47,"value":457},": If an agent fails or times out, retry once. Transient failures\nhappen.",{"type":41,"tag":282,"props":459,"children":461},{"id":460},"_3-clean-the-solution-context-isolation-the-1-lever",[462],{"type":47,"value":463},"3. Clean the solution (context isolation — the #1 lever)",{"type":41,"tag":117,"props":465,"children":466},{},[467],{"type":47,"value":468},"The thinking trace biases the verifier toward agreement — a long chain of\nreasoning reads as supporting evidence even when the conclusion is wrong. Before\nany verification, strip:",{"type":41,"tag":391,"props":470,"children":471},{},[472,477,482],{"type":41,"tag":61,"props":473,"children":474},{},[475],{"type":47,"value":476},"All thinking-block content",{"type":41,"tag":61,"props":478,"children":479},{},[480],{"type":47,"value":481},"All \"Let me try...\" \u002F \"Actually wait...\" \u002F \"Hmm\" prose",{"type":41,"tag":61,"props":483,"children":484},{},[485],{"type":47,"value":486},"All false starts and backtracking",{"type":41,"tag":117,"props":488,"children":489},{},[490],{"type":47,"value":491},"What remains: problem statement + clean final argument only.",{"type":41,"tag":117,"props":493,"children":494},{},[495,497,502,504,509,510,515],{"type":47,"value":496},"Extract only the ",{"type":41,"tag":65,"props":498,"children":499},{},[500],{"type":47,"value":501},"Method",{"type":47,"value":503}," + ",{"type":41,"tag":65,"props":505,"children":506},{},[507],{"type":47,"value":508},"Proof",{"type":47,"value":503},{"type":41,"tag":65,"props":511,"children":512},{},[513],{"type":47,"value":514},"Answer",{"type":47,"value":516}," sections from each solver's\noutput. The verifier never sees how the solver got there.",{"type":41,"tag":282,"props":518,"children":520},{"id":519},"_4-adversarial-verify-fresh-context-pattern-armed",[521],{"type":47,"value":522},"4. Adversarial verify (fresh context, pattern-armed)",{"type":41,"tag":117,"props":524,"children":525},{},[526,528,533,535],{"type":47,"value":527},"For each cleaned solution, launch a fresh verifier agent. ",{"type":41,"tag":65,"props":529,"children":530},{},[531],{"type":47,"value":532},"Fresh context",{"type":47,"value":534},": it\nsees only (problem statement + cleaned solution). ",{"type":41,"tag":65,"props":536,"children":537},{},[538],{"type":47,"value":539},"No tools.",{"type":41,"tag":117,"props":541,"children":542},{},[543,545,551],{"type":47,"value":544},"The verifier's job is to ATTACK, not grade. Load\n",{"type":41,"tag":257,"props":546,"children":548},{"className":547},[],[549],{"type":47,"value":550},"references\u002Fadversarial_prompts.md",{"type":47,"value":552}," for the prompts. The key patterns it runs:",{"type":41,"tag":137,"props":554,"children":555},{},[556,572],{"type":41,"tag":141,"props":557,"children":558},{},[559],{"type":41,"tag":145,"props":560,"children":561},{},[562,567],{"type":41,"tag":149,"props":563,"children":564},{},[565],{"type":47,"value":566},"Pattern",{"type":41,"tag":149,"props":568,"children":569},{},[570],{"type":47,"value":571},"The check",{"type":41,"tag":165,"props":573,"children":574},{},[575,591,607,623,647],{"type":41,"tag":145,"props":576,"children":577},{},[578,586],{"type":41,"tag":172,"props":579,"children":580},{},[581],{"type":41,"tag":65,"props":582,"children":583},{},[584],{"type":47,"value":585},"#4",{"type":41,"tag":172,"props":587,"children":588},{},[589],{"type":47,"value":590},"Does this theorem specialize to a famous object (ζ, quadratic reciprocity, etc.) and prove something open about it? → gap",{"type":41,"tag":145,"props":592,"children":593},{},[594,602],{"type":41,"tag":172,"props":595,"children":596},{},[597],{"type":41,"tag":65,"props":598,"children":599},{},[600],{"type":47,"value":601},"#18",{"type":41,"tag":172,"props":603,"children":604},{},[605],{"type":47,"value":606},"Substitute the proof's own intermediate identities into any \"remaining gap.\" Recover the original claim? → tautological",{"type":41,"tag":145,"props":608,"children":609},{},[610,618],{"type":41,"tag":172,"props":611,"children":612},{},[613],{"type":41,"tag":65,"props":614,"children":615},{},[616],{"type":47,"value":617},"#40",{"type":41,"tag":172,"props":619,"children":620},{},[621],{"type":47,"value":622},"Is any step a \"one-line lemma\"? Extract the GENERAL form. Find a 2×2 counterexample. If the general form is false, find what special structure saves THIS instance",{"type":41,"tag":145,"props":624,"children":625},{},[626,634],{"type":41,"tag":172,"props":627,"children":628},{},[629],{"type":41,"tag":65,"props":630,"children":631},{},[632],{"type":47,"value":633},"#5",{"type":41,"tag":172,"props":635,"children":636},{},[637,639,645],{"type":47,"value":638},"For each invoked theorem: re-check hypotheses FROM SCRATCH. \"Continuous on ",{"type":41,"tag":640,"props":641,"children":642},"span",{},[643],{"type":47,"value":644},"0,1",{"type":47,"value":646},"\" ≠ \"continuous on ℝ\"",{"type":41,"tag":145,"props":648,"children":649},{},[650,658],{"type":41,"tag":172,"props":651,"children":652},{},[653],{"type":41,"tag":65,"props":654,"children":655},{},[656],{"type":47,"value":657},"#6",{"type":41,"tag":172,"props":659,"children":660},{},[661],{"type":47,"value":662},"Any infinite sum \"bounded\" via a regularized value? Check the boundary — if there's a pole there, the sum diverges",{"type":41,"tag":117,"props":664,"children":665},{},[666,668],{"type":47,"value":667},"Full pattern list: ",{"type":41,"tag":257,"props":669,"children":671},{"className":670},[],[672],{"type":47,"value":673},"references\u002Fverifier_patterns.md",{"type":41,"tag":117,"props":675,"children":676},{},[677],{"type":47,"value":678},"Verifier returns:",{"type":41,"tag":361,"props":680,"children":683},{"className":681,"code":682,"language":47},[364],"**Verdict**: HOLDS | HOLE FOUND | UNCLEAR\n\n**If HOLE FOUND**:\n- Location: [quote the problematic step]\n- Pattern: [which check fired, or \"other\"]\n- Why it breaks: [specific]\n- Fixable?: [yes with X \u002F no, fundamental]\n",[684],{"type":41,"tag":257,"props":685,"children":686},{"__ignoreMap":369},[687],{"type":47,"value":682},{"type":41,"tag":282,"props":689,"children":691},{"id":690},"_5-rank-and-vote-verify-asymmetric-early-exit",[692],{"type":47,"value":693},"5. Rank and vote-verify (asymmetric + early exit)",{"type":41,"tag":117,"props":695,"children":696},{},[697],{"type":47,"value":698},"Rank solutions by (verdict, verifier confidence). Take the top one. Run up to 5\nfresh verifier agents.",{"type":41,"tag":117,"props":700,"children":701},{},[702,707],{"type":41,"tag":65,"props":703,"children":704},{},[705],{"type":47,"value":706},"Asymmetric thresholds",{"type":47,"value":708},": 4 HOLDS to confirm, 2 HOLE FOUND to refute. Why\nasymmetric: one flaky verifier shouldn't kill a correct proof; but two\nindependent dissents is a real signal.",{"type":41,"tag":117,"props":710,"children":711},{},[712,717],{"type":41,"tag":65,"props":713,"children":714},{},[715],{"type":47,"value":716},"Pigeonhole early exit",{"type":47,"value":718},": stop launching verifiers once the outcome is decided.",{"type":41,"tag":391,"props":720,"children":721},{},[722,727,732],{"type":41,"tag":61,"props":723,"children":724},{},[725],{"type":47,"value":726},"2 say HOLE FOUND → refuted, stop (save the remaining 3 calls)",{"type":41,"tag":61,"props":728,"children":729},{},[730],{"type":47,"value":731},"4 say HOLDS → confirmed, stop (save the 5th)",{"type":41,"tag":61,"props":733,"children":734},{},[735],{"type":47,"value":736},"After 3 verifiers: if 2 HOLDS + 1 HOLE, launch 2 more (outcome undecided). If\n3 HOLDS + 0 HOLE, launch 1 more (could still hit 4-1).",{"type":41,"tag":117,"props":738,"children":739},{},[740,745],{"type":41,"tag":65,"props":741,"children":742},{},[743],{"type":47,"value":744},"Dual context-isolation",{"type":47,"value":746},": each verifier is blind to (a) the solver's thinking\ntrace — already stripped in step 3 — AND (b) other verifiers' verdicts. Each\nverifier thinks it's the first. No \"3 agents already confirmed this\" social\nproof.",{"type":41,"tag":117,"props":748,"children":749},{},[750,755],{"type":41,"tag":65,"props":751,"children":752},{},[753],{"type":47,"value":754},"A solver cannot verify its own solution.",{"type":47,"value":756}," Different agent, fresh context.",{"type":41,"tag":282,"props":758,"children":760},{"id":759},"_5b-when-one-case-wont-close-step-back-before-grinding",[761],{"type":47,"value":762},"5b. When one case won't close — step back before grinding",{"type":41,"tag":117,"props":764,"children":765},{},[766,768],{"type":47,"value":767},"If a proof splits into cases and one case proves easily but the other resists:\n",{"type":41,"tag":65,"props":769,"children":770},{},[771],{"type":47,"value":772},"before grinding through the hard case, ask whether there's a route that makes\nthe split disappear.",{"type":41,"tag":117,"props":774,"children":775},{},[776,778,784],{"type":47,"value":777},"The pattern that saves you: the hard case's very hypothesis often implies\nsomething strong about an ",{"type":41,"tag":779,"props":780,"children":781},"em",{},[782],{"type":47,"value":783},"intermediate object",{"type":47,"value":785}," you haven't looked at. Use that\nimplication directly instead of the original chain.",{"type":41,"tag":117,"props":787,"children":788},{},[789,791,796],{"type":47,"value":790},"Concrete shape: proving f(n) ≤ cn for a constrained function f, with a case\nsplit on a prime p dividing f(n). One branch closes by index arguments in\n(ℤ\u002Fp^e)*. The other branch resists — same group structure, but the arithmetic\ndoesn't contradict. The fix: the hypothesis \"p | f(n)\" plugged back into the\ngoverning equation implies ",{"type":41,"tag":65,"props":792,"children":793},{},[794],{"type":47,"value":795},"f(p) = p itself",{"type":47,"value":797},". Once you have that, a\nFermat+Dirichlet argument kills both branches in three lines. The case split was\na detour — it was splitting on a variable that, under the hypothesis, takes a\nknown value.",{"type":41,"tag":117,"props":799,"children":800},{},[801],{"type":47,"value":802},"Check when stuck on case B:",{"type":41,"tag":391,"props":804,"children":805},{},[806,818,823],{"type":41,"tag":61,"props":807,"children":808},{},[809,811,816],{"type":47,"value":810},"What does case B's hypothesis imply about f at ",{"type":41,"tag":779,"props":812,"children":813},{},[814],{"type":47,"value":815},"other",{"type":47,"value":817}," inputs?",{"type":41,"tag":61,"props":819,"children":820},{},[821],{"type":47,"value":822},"Is there a different pair (a,b) to plug into the governing equation?",{"type":41,"tag":61,"props":824,"children":825},{},[826],{"type":47,"value":827},"Are you proving too much? (A cleaner contradiction needs less machinery.)",{"type":41,"tag":117,"props":829,"children":830},{},[831],{"type":47,"value":832},"This is also a presentation-pass win: the split-free proof is shorter AND more\ngeneral.",{"type":41,"tag":282,"props":834,"children":836},{"id":835},"_6-revise-if-needed",[837],{"type":47,"value":838},"6. Revise (if needed)",{"type":41,"tag":117,"props":840,"children":841},{},[842],{"type":47,"value":843},"If verification finds a hole: launch a reviser agent. It gets (cleaned\nsolution + verifier's hole report). STILL no access to the original thinking —\nthe reviser works from the hole, not by rereading how you got there.",{"type":41,"tag":361,"props":845,"children":848},{"className":846,"code":847,"language":47},[364],"A verifier found this issue in the proof:\n[hole report]\n\nFix the proof. If the hole is fundamental (the approach doesn't work), say so and return **Verdict: no confident solution** with what partial progress remains.\n\nFor any step you cannot fully close, mark it inline: [GAP: specific description of what remains]. Gaps in the proof text, not in a separate list — they're greppable and the next reviser knows exactly where to look.\n",[849],{"type":41,"tag":257,"props":850,"children":851},{"__ignoreMap":369},[852],{"type":47,"value":847},{"type":41,"tag":117,"props":854,"children":855},{},[856],{"type":47,"value":857},"Up to 3 revise cycles. Then re-run the vote on the revised proof.",{"type":41,"tag":117,"props":859,"children":860},{},[861,866,868,873],{"type":41,"tag":65,"props":862,"children":863},{},[864],{"type":47,"value":865},"If pattern #40 fired",{"type":47,"value":867}," (one-line-proof-too-clean), the reviser gets a stronger\nbrief — the Adversarial Brief template from ",{"type":41,"tag":257,"props":869,"children":871},{"className":870},[],[872],{"type":47,"value":550},{"type":47,"value":874},"\n§7. It forces a binary: \"the general lemma is obviously false (here's a 2×2\ncounterexample) — so either find what's special about THIS case, or find where\nthe proof breaks.\" Can't return \"looks fine.\"",{"type":41,"tag":282,"props":876,"children":878},{"id":877},"_6c-deep-mode-when-tight-budget-abstains",[879],{"type":47,"value":880},"6c. Deep mode (when tight-budget abstains)",{"type":41,"tag":117,"props":882,"children":883},{},[884],{"type":47,"value":885},"The standard workflow is tight-budget: 8 solvers, ~15 min, pure reasoning. When\nit abstains, the problem may need more time, not more capability.",{"type":41,"tag":117,"props":887,"children":888},{},[889,894],{"type":41,"tag":65,"props":890,"children":891},{},[892],{"type":47,"value":893},"Deep mode",{"type":47,"value":895}," is a single focused agent with:",{"type":41,"tag":391,"props":897,"children":898},{},[899,909,919],{"type":41,"tag":61,"props":900,"children":901},{},[902,907],{"type":41,"tag":65,"props":903,"children":904},{},[905],{"type":47,"value":906},"Unlimited time",{"type":47,"value":908}," — no wall-clock pressure",{"type":41,"tag":61,"props":910,"children":911},{},[912,917],{"type":41,"tag":65,"props":913,"children":914},{},[915],{"type":47,"value":916},"Targeted computation allowed",{"type":47,"value":918}," — modular arithmetic checks, small-case\nenumeration, symbolic verification of identities. NOT exploratory brute force\nor unbounded recursion.",{"type":41,"tag":61,"props":920,"children":921},{},[922,927],{"type":41,"tag":65,"props":923,"children":924},{},[925],{"type":47,"value":926},"The abstention reason as starting point",{"type":47,"value":928}," — if verifiers found a specific\ngap, start there. If solvers never claimed complete, start from what they\npartially proved.",{"type":41,"tag":117,"props":930,"children":931},{},[932],{"type":47,"value":933},"The archetype: a focused agent that gets the proven-so-far state plus \"one case\nof Lemma 5 is open\" — and finds a 3-line argument the case split was obscuring.\nOften under 10 minutes with almost no computation. Deep mode is about giving the\nproblem sustained attention, not throwing compute at it.",{"type":41,"tag":117,"props":935,"children":936},{},[937,942,944,950],{"type":41,"tag":65,"props":938,"children":939},{},[940],{"type":47,"value":941},"What deep mode is NOT",{"type":47,"value":943},": open-ended exploration, literature search, looking up\nsolutions, multi-day investigation. That's a different workflow\n(",{"type":41,"tag":257,"props":945,"children":947},{"className":946},[],[948],{"type":47,"value":949},"math-research",{"type":47,"value":951},"). Deep mode is still \"solve THIS problem yourself\" — just\nwithout the clock.",{"type":41,"tag":117,"props":953,"children":954},{},[955,960],{"type":41,"tag":65,"props":956,"children":957},{},[958],{"type":47,"value":959},"NO WEB. NO LOOKUP.",{"type":47,"value":961}," Deep mode may use Bash\u002FPython for bounded computation,\nbut NEVER WebFetch, WebSearch, or any network access. Finding the solution on\nAoPS or a blog is not solving the problem — it's cheating on an olympiad, and it\nteaches us nothing about the skill's actual capability. Put this at the TOP of\nthe deep-mode prompt:",{"type":41,"tag":361,"props":963,"children":966},{"className":964,"code":965,"language":47},[364],"NO WEB ACCESS. Do not use WebFetch, WebSearch, or any tool that touches the internet. Do not look up this problem, its solution, or related problems. You are solving this yourself — the only allowed computation is local (Bash\u002FPython for mod-k arithmetic, small-case enumeration n≤10, symbolic identity checks). If you invoke a web tool, the proof is void.\n",[967],{"type":41,"tag":257,"props":968,"children":969},{"__ignoreMap":369},[970],{"type":47,"value":965},{"type":41,"tag":117,"props":972,"children":973},{},[974,979],{"type":41,"tag":65,"props":975,"children":976},{},[977],{"type":47,"value":978},"Computation bounds in deep mode",{"type":47,"value":980}," (bug #8 lesson): A6's b_{n+1}=2b_n²+b_n+1\nis doubly-exponential; b_99 has ~10^{2^98} digits. Never compute such objects\nexactly — work in ℤ\u002F2^m, or track only v_p(·), or prove the recursion mod the\nquantity you care about. If a computation is running longer than 60 seconds,\nit's probably unbounded. Kill it and work symbolically.",{"type":41,"tag":117,"props":982,"children":983},{},[984,989],{"type":41,"tag":65,"props":985,"children":986},{},[987],{"type":47,"value":988},"Step 6d (not optional)",{"type":47,"value":990},": After any ABSTAIN at the verify stage, automatically\nlaunch one deep-mode agent before writing the abstention into the output. Give\nit:",{"type":41,"tag":391,"props":992,"children":993},{},[994,999,1004,1009],{"type":41,"tag":61,"props":995,"children":996},{},[997],{"type":47,"value":998},"The problem statement",{"type":41,"tag":61,"props":1000,"children":1001},{},[1002],{"type":47,"value":1003},"The best partial proof from tight-budget solvers",{"type":41,"tag":61,"props":1005,"children":1006},{},[1007],{"type":47,"value":1008},"The verifier gap descriptions (what specifically didn't close)",{"type":41,"tag":61,"props":1010,"children":1011},{},[1012],{"type":47,"value":1013},"The instruction: \"NO WEB ACCESS — do not look up this problem or its solution.\nBounded local computation allowed (mod 2^k, small cases n≤10, symbolic\nidentity checks via Bash\u002FPython only). 60-second computation limit. If n≤10\nbrute force reveals a pattern the tight-budget solvers missed, that pattern IS\nthe proof structure.\"",{"type":41,"tag":117,"props":1015,"children":1016},{},[1017],{"type":47,"value":1018},"The deep agent may find the construction the pure-reasoning solvers couldn't\nsee. If it also abstains, THEN write the abstention. Do not skip this step —\nproblems with √n or log n answers are often invisible to pure reasoning because\nthe optimal structure is the asymmetric one.",{"type":41,"tag":117,"props":1020,"children":1021},{},[1022,1027],{"type":41,"tag":65,"props":1023,"children":1024},{},[1025],{"type":47,"value":1026},"Orchestrator self-restraint",{"type":47,"value":1028},": The orchestrator itself must not web-search the\nproblem \"to help\" the deep agent. If you're tempted to Fetch an AoPS thread\n\"just to check the answer,\" don't — that contaminates the skill's output and\nmisrepresents its capability.",{"type":41,"tag":282,"props":1030,"children":1032},{"id":1031},"_7-calibrated-abstention",[1033],{"type":47,"value":1034},"7. Calibrated abstention",{"type":41,"tag":117,"props":1036,"children":1037},{},[1038,1040],{"type":47,"value":1039},"If 3 revise cycles all fail: ",{"type":41,"tag":65,"props":1041,"children":1042},{},[1043],{"type":47,"value":1044},"stop and admit it.",{"type":41,"tag":361,"props":1046,"children":1049},{"className":1047,"code":1048,"language":47},[364],"**Verdict**: no confident solution\n\n**What was tried**: [approaches]\n**What WAS proven**: [any lemma or partial result that survived verification]\n**Where it breaks**: [the unfixed hole]\n",[1050],{"type":41,"tag":257,"props":1051,"children":1052},{"__ignoreMap":369},[1053],{"type":47,"value":1048},{"type":41,"tag":117,"props":1055,"children":1056},{},[1057],{"type":47,"value":1058},"Do NOT guess. A wrong confident answer is worse than an honest \"couldn't solve\nit.\" The metric that matters is CONDITIONAL accuracy — when you say \"solved,\"\nare you right?",{"type":41,"tag":282,"props":1060,"children":1062},{"id":1061},"_8-presentation-pass-after-correctness-is-established",[1063],{"type":47,"value":1064},"8. Presentation pass (after correctness is established)",{"type":41,"tag":117,"props":1066,"children":1067},{},[1068],{"type":47,"value":1069},"A VERIFIED-CORRECT proof is often not a BEAUTIFUL proof. The order you\ndiscovered it is rarely the best order to present it. Launch a fresh\npresentation agent with the verified proof.",{"type":41,"tag":117,"props":1071,"children":1072},{},[1073,1075,1081],{"type":47,"value":1074},"Load ",{"type":41,"tag":257,"props":1076,"children":1078},{"className":1077},[],[1079],{"type":47,"value":1080},"references\u002Fpresentation_prompts.md",{"type":47,"value":1082},". The agent asks:",{"type":41,"tag":391,"props":1084,"children":1085},{},[1086,1091,1096,1101],{"type":41,"tag":61,"props":1087,"children":1088},{},[1089],{"type":47,"value":1090},"What's the simplest way to say this?",{"type":41,"tag":61,"props":1092,"children":1093},{},[1094],{"type":47,"value":1095},"Which lemmas should be inlined? Which deserve to stand alone?",{"type":41,"tag":61,"props":1097,"children":1098},{},[1099],{"type":47,"value":1100},"Is anything OVERKILL? (constructing a double exponential when linear suffices)",{"type":41,"tag":61,"props":1102,"children":1103},{},[1104],{"type":47,"value":1105},"Now that we know the answer, is there a 3-line hindsight proof?",{"type":41,"tag":117,"props":1107,"children":1108},{},[1109,1111,1117,1119,1125,1127,1133],{"type":47,"value":1110},"Output: LaTeX-formatted proof. If ",{"type":41,"tag":257,"props":1112,"children":1114},{"className":1113},[],[1115],{"type":47,"value":1116},"pdflatex",{"type":47,"value":1118}," is available\n(",{"type":41,"tag":257,"props":1120,"children":1122},{"className":1121},[],[1123],{"type":47,"value":1124},"scripts\u002Fcheck_latex.sh",{"type":47,"value":1126}," returns 0), also compile to PDF via\n",{"type":41,"tag":257,"props":1128,"children":1130},{"className":1129},[],[1131],{"type":47,"value":1132},"scripts\u002Fcompile_pdf.sh",{"type":47,"value":1134},".",{"type":41,"tag":113,"props":1136,"children":1137},{},[],{"type":41,"tag":50,"props":1139,"children":1141},{"id":1140},"model-tier-defaults",[1142],{"type":47,"value":1143},"Model tier defaults",{"type":41,"tag":117,"props":1145,"children":1146},{},[1147,1149,1155],{"type":47,"value":1148},"Read ",{"type":41,"tag":257,"props":1150,"children":1152},{"className":1151},[],[1153],{"type":47,"value":1154},"references\u002Fmodel_tier_defaults.md",{"type":47,"value":1156}," for full details. Summary:",{"type":41,"tag":137,"props":1158,"children":1159},{},[1160,1191],{"type":41,"tag":141,"props":1161,"children":1162},{},[1163],{"type":41,"tag":145,"props":1164,"children":1165},{},[1166,1171,1176,1181,1186],{"type":41,"tag":149,"props":1167,"children":1168},{},[1169],{"type":47,"value":1170},"Model",{"type":41,"tag":149,"props":1172,"children":1173},{},[1174],{"type":47,"value":1175},"Solvers",{"type":41,"tag":149,"props":1177,"children":1178},{},[1179],{"type":47,"value":1180},"Verify passes",{"type":41,"tag":149,"props":1182,"children":1183},{},[1184],{"type":47,"value":1185},"Abstain after",{"type":41,"tag":149,"props":1187,"children":1188},{},[1189],{"type":47,"value":1190},"Presentation",{"type":41,"tag":165,"props":1192,"children":1193},{},[1194,1222,1250],{"type":41,"tag":145,"props":1195,"children":1196},{},[1197,1202,1207,1212,1217],{"type":41,"tag":172,"props":1198,"children":1199},{},[1200],{"type":47,"value":1201},"Haiku",{"type":41,"tag":172,"props":1203,"children":1204},{},[1205],{"type":47,"value":1206},"8",{"type":41,"tag":172,"props":1208,"children":1209},{},[1210],{"type":47,"value":1211},"3",{"type":41,"tag":172,"props":1213,"children":1214},{},[1215],{"type":47,"value":1216},"2 revise fails",{"type":41,"tag":172,"props":1218,"children":1219},{},[1220],{"type":47,"value":1221},"skip",{"type":41,"tag":145,"props":1223,"children":1224},{},[1225,1230,1235,1240,1245],{"type":41,"tag":172,"props":1226,"children":1227},{},[1228],{"type":47,"value":1229},"Sonnet",{"type":41,"tag":172,"props":1231,"children":1232},{},[1233],{"type":47,"value":1234},"4",{"type":41,"tag":172,"props":1236,"children":1237},{},[1238],{"type":47,"value":1239},"5",{"type":41,"tag":172,"props":1241,"children":1242},{},[1243],{"type":47,"value":1244},"3 revise fails",{"type":41,"tag":172,"props":1246,"children":1247},{},[1248],{"type":47,"value":1249},"yes",{"type":41,"tag":145,"props":1251,"children":1252},{},[1253,1258,1262,1267,1272],{"type":41,"tag":172,"props":1254,"children":1255},{},[1256],{"type":47,"value":1257},"Opus",{"type":41,"tag":172,"props":1259,"children":1260},{},[1261],{"type":47,"value":1211},{"type":41,"tag":172,"props":1263,"children":1264},{},[1265],{"type":47,"value":1266},"5 + full pattern sweep",{"type":41,"tag":172,"props":1268,"children":1269},{},[1270],{"type":47,"value":1271},"4 revise fails",{"type":41,"tag":172,"props":1273,"children":1274},{},[1275],{"type":47,"value":1276},"2 drafts, pick cleaner",{"type":41,"tag":117,"props":1278,"children":1279},{},[1280],{"type":47,"value":1281},"Weaker models: more parallel attempts, faster abstention. Stronger models:\ndeeper verification, more presentation effort.",{"type":41,"tag":113,"props":1283,"children":1284},{},[],{"type":41,"tag":50,"props":1286,"children":1288},{"id":1287},"for-numeric-answer-problems-aime-style",[1289],{"type":47,"value":1290},"For numeric-answer problems (AIME-style)",{"type":41,"tag":117,"props":1292,"children":1293},{},[1294],{"type":47,"value":1295},"Skip the proof machinery. Run 5-7 solvers with varied approaches, take majority\nvote on the numeric answer. If no majority: verify the top 2 candidates by\nsubstitution.",{"type":41,"tag":113,"props":1297,"children":1298},{},[],{"type":41,"tag":50,"props":1300,"children":1302},{"id":1301},"key-references",[1303],{"type":47,"value":1304},"Key references",{"type":41,"tag":391,"props":1306,"children":1307},{},[1308,1318,1328,1338],{"type":41,"tag":61,"props":1309,"children":1310},{},[1311,1316],{"type":41,"tag":257,"props":1312,"children":1314},{"className":1313},[],[1315],{"type":47,"value":673},{"type":47,"value":1317}," — the 12 adversarial checks",{"type":41,"tag":61,"props":1319,"children":1320},{},[1321,1326],{"type":41,"tag":257,"props":1322,"children":1324},{"className":1323},[],[1325],{"type":47,"value":550},{"type":47,"value":1327}," — ready-to-use verifier prompts",{"type":41,"tag":61,"props":1329,"children":1330},{},[1331,1336],{"type":41,"tag":257,"props":1332,"children":1334},{"className":1333},[],[1335],{"type":47,"value":1080},{"type":47,"value":1337}," — beautification prompts + LaTeX template",{"type":41,"tag":61,"props":1339,"children":1340},{},[1341,1346],{"type":41,"tag":257,"props":1342,"children":1344},{"className":1343},[],[1345],{"type":47,"value":1154},{"type":47,"value":1347}," — per-model configuration",{"type":41,"tag":113,"props":1349,"children":1350},{},[],{"type":41,"tag":50,"props":1352,"children":1354},{"id":1353},"what-makes-this-different-from-generic-verify-and-refine",[1355],{"type":47,"value":1356},"What makes this different from generic verify-and-refine",{"type":41,"tag":57,"props":1358,"children":1359},{},[1360,1370,1380,1390,1400,1410],{"type":41,"tag":61,"props":1361,"children":1362},{},[1363,1368],{"type":41,"tag":65,"props":1364,"children":1365},{},[1366],{"type":47,"value":1367},"Dual context isolation",{"type":47,"value":1369},": verifier is blind to (a) the solver's thinking\ntrace — which biases toward agreement — and (b) other verifiers' verdicts —\nsocial proof also biases. Each verifier thinks it's first.",{"type":41,"tag":61,"props":1371,"children":1372},{},[1373,1378],{"type":41,"tag":65,"props":1374,"children":1375},{},[1376],{"type":47,"value":1377},"Pattern-specific attacks",{"type":47,"value":1379},": not \"is this correct?\" but \"does this make the\n#40 mistake? the #4 mistake?\" Specific beats generic. The 7-category\nrefutation taxonomy gives the verifier a checklist.",{"type":41,"tag":61,"props":1381,"children":1382},{},[1383,1388],{"type":41,"tag":65,"props":1384,"children":1385},{},[1386],{"type":47,"value":1387},"Asymmetric vote + pigeonhole exit",{"type":47,"value":1389},": 4-to-confirm, 2-to-refute. One flaky\nverifier doesn't kill a correct proof; two dissents does. Stop launching\nverifiers once the outcome is decided — saves ~30% of verification cost on\nclear cases.",{"type":41,"tag":61,"props":1391,"children":1392},{},[1393,1398],{"type":41,"tag":65,"props":1394,"children":1395},{},[1396],{"type":47,"value":1397},"Specification-gaming check first",{"type":47,"value":1399},": explicitly asks \"is this the intended\ninterpretation?\" before solving. The #1 failure mode in prior work (50\u002F63\n\"correct\" answers solved the wrong reading).",{"type":41,"tag":61,"props":1401,"children":1402},{},[1403,1408],{"type":41,"tag":65,"props":1404,"children":1405},{},[1406],{"type":47,"value":1407},"Calibrated abstention",{"type":47,"value":1409},": will say \"no confident solution\" with partial\nresults. Optimizes conditional accuracy, not coverage.",{"type":41,"tag":61,"props":1411,"children":1412},{},[1413,1418],{"type":41,"tag":65,"props":1414,"children":1415},{},[1416],{"type":47,"value":1417},"Presentation pass",{"type":47,"value":1419},": correctness and elegance are separate steps. The\npresentation agent gets the VERIFIED proof and finds the cleanest way to say\nit.",{"items":1421,"total":1527},[1422,1438,1455,1470,1483,1496,1515],{"slug":1423,"name":1423,"fn":1424,"description":1425,"org":1426,"tags":1427,"stars":20,"repoUrl":21,"updatedAt":1437},"access","manage iMessage channel access and permissions","Manage iMessage channel access — approve pairings, edit allowlists, set DM\u002Fgroup policy. Use when the user asks to pair, approve someone, check who's allowed, or change policy for the iMessage channel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1428,1431,1434],{"name":1429,"slug":1430,"type":16},"Access Control","access-control",{"name":1432,"slug":1433,"type":16},"iMessage","imessage",{"name":1435,"slug":1436,"type":16},"Messaging","messaging","2026-04-12T04:54:55.917969",{"slug":1439,"name":1439,"fn":1440,"description":1441,"org":1442,"tags":1443,"stars":20,"repoUrl":21,"updatedAt":1454},"agent-development","develop and configure AI agents","This skill should be used when the user asks to \"create an agent\", \"add an agent\", \"write a subagent\", \"agent frontmatter\", \"when to use description\", \"agent examples\", \"agent tools\", \"agent colors\", \"autonomous agent\", or needs guidance on agent structure, system prompts, triggering conditions, or agent development best practices for Claude Code plugins.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1444,1447,1448,1451],{"name":1445,"slug":1446,"type":16},"Agents","agents",{"name":9,"slug":8,"type":16},{"name":1449,"slug":1450,"type":16},"Documentation","documentation",{"name":1452,"slug":1453,"type":16},"Plugin Development","plugin-development","2026-04-10T04:55:41.251084",{"slug":1456,"name":1456,"fn":1457,"description":1458,"org":1459,"tags":1460,"stars":20,"repoUrl":21,"updatedAt":1469},"build-mcp-app","build MCP apps with interactive UI","This skill should be used when the user wants to build an \"MCP app\", add \"interactive UI\" or \"widgets\" to an MCP server, \"render components in chat\", build \"MCP UI resources\", make a tool that shows a \"form\", \"picker\", \"dashboard\" or \"confirmation dialog\" inline in the conversation, or mentions \"apps SDK\" in the context of MCP. Use AFTER the build-mcp-server skill has settled the deployment model, or when the user already knows they want UI widgets.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1461,1462,1464,1466],{"name":1445,"slug":1446,"type":16},{"name":1463,"slug":26,"type":16},"Claude Code",{"name":1465,"slug":27,"type":16},"MCP",{"name":1467,"slug":1468,"type":16},"UI Components","ui-components","2026-04-06T17:59:32.421865",{"slug":1471,"name":1471,"fn":1472,"description":1473,"org":1474,"tags":1475,"stars":20,"repoUrl":21,"updatedAt":1482},"build-mcp-server","build MCP servers","This skill should be used when the user asks to \"build an MCP server\", \"create an MCP\", \"make an MCP integration\", \"wrap an API for Claude\", \"expose tools to Claude\", \"make an MCP app\", or discusses building something with the Model Context Protocol. It is the entry point for MCP server development — it interrogates the user about their use case, determines the right deployment model (remote HTTP, MCPB, local stdio), picks a tool-design pattern, and hands off to specialized skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1476,1477,1480,1481],{"name":1445,"slug":1446,"type":16},{"name":1478,"slug":1479,"type":16},"API Development","api-development",{"name":1463,"slug":26,"type":16},{"name":1465,"slug":27,"type":16},"2026-04-06T17:59:33.744601",{"slug":1484,"name":1484,"fn":1485,"description":1486,"org":1487,"tags":1488,"stars":20,"repoUrl":21,"updatedAt":1495},"build-mcpb","package and distribute MCP servers","This skill should be used when the user wants to \"package an MCP server\", \"bundle an MCP\", \"make an MCPB\", \"ship a local MCP server\", \"distribute a local MCP\", discusses \".mcpb files\", mentions bundling a Node or Python runtime with their MCP server, or needs an MCP server that interacts with the local filesystem, desktop apps, or OS and must be installable without the user having Node\u002FPython set up.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1489,1490,1491,1492],{"name":1445,"slug":1446,"type":16},{"name":1463,"slug":26,"type":16},{"name":1465,"slug":27,"type":16},{"name":1493,"slug":1494,"type":16},"Packaging","packaging","2026-04-06T17:59:31.159961",{"slug":1497,"name":1497,"fn":1498,"description":1499,"org":1500,"tags":1501,"stars":20,"repoUrl":21,"updatedAt":1514},"cardputer-buddy","develop MicroPython apps for Cardputer","Iterate on the Cardputer-Adv MicroPython app bundle (Claude Buddy, Snake, Hello) after the device is already provisioned via m5-onboard. Use when the user wants to add a new app, push a single changed .py without re-flashing, watch device serial logs, or run a one-shot REPL command. Trigger on \"add an app\", \"push to the cardputer\", \"tail the device\", \"run on the device\", or follow-up work after \u002Fmaker-setup.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1502,1505,1508,1511],{"name":1503,"slug":1504,"type":16},"Hardware","hardware",{"name":1506,"slug":1507,"type":16},"M5Stack","m5stack",{"name":1509,"slug":1510,"type":16},"MicroPython","micropython",{"name":1512,"slug":1513,"type":16},"Python","python","2026-05-06T05:39:00.134385",{"slug":1516,"name":1516,"fn":1517,"description":1518,"org":1519,"tags":1520,"stars":20,"repoUrl":21,"updatedAt":1526},"claude-automation-recommender","recommend Claude Code automations","Analyze a codebase and recommend Claude Code automations (hooks, subagents, skills, plugins, MCP servers). Use when user asks for automation recommendations, wants to optimize their Claude Code setup, mentions improving Claude Code workflows, asks how to first set up Claude Code for a project, or wants to know what Claude Code features they should use.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1521,1524,1525],{"name":1522,"slug":1523,"type":16},"Agent Context","agent-context",{"name":1463,"slug":26,"type":16},{"name":1465,"slug":27,"type":16},"2026-04-06T18:00:34.049624",25,{"items":1529,"total":1705},[1530,1551,1565,1577,1594,1605,1626,1644,1658,1668,1676,1689],{"slug":1531,"name":1531,"fn":1532,"description":1533,"org":1534,"tags":1535,"stars":1548,"repoUrl":1549,"updatedAt":1550},"algorithmic-art","create algorithmic art with p5.js","Creating algorithmic art using p5.js with seeded randomness and interactive parameter exploration. Use this when users request creating art using code, generative art, algorithmic art, flow fields, or particle systems. Create original algorithmic art rather than copying existing artists' work to avoid copyright violations.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1536,1539,1542,1545],{"name":1537,"slug":1538,"type":16},"Creative","creative",{"name":1540,"slug":1541,"type":16},"Design","design",{"name":1543,"slug":1544,"type":16},"Generative Art","generative-art",{"name":1546,"slug":1547,"type":16},"JavaScript","javascript",161831,"https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fskills","2026-04-06T17:56:15.455818",{"slug":1552,"name":1552,"fn":1553,"description":1554,"org":1555,"tags":1556,"stars":1548,"repoUrl":1549,"updatedAt":1564},"brand-guidelines","apply Anthropic brand colors and typography","Applies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1557,1560,1561],{"name":1558,"slug":1559,"type":16},"Branding","branding",{"name":1540,"slug":1541,"type":16},{"name":1562,"slug":1563,"type":16},"Typography","typography","2026-04-06T17:56:05.042852",{"slug":1566,"name":1566,"fn":1567,"description":1568,"org":1569,"tags":1570,"stars":1548,"repoUrl":1549,"updatedAt":1576},"canvas-design","create posters and visual art as PNG or PDF","Create beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1571,1572,1573],{"name":1537,"slug":1538,"type":16},{"name":1540,"slug":1541,"type":16},{"name":1574,"slug":1575,"type":16},"PDF","pdf","2026-04-06T17:56:03.794732",{"slug":1578,"name":1578,"fn":1579,"description":1580,"org":1581,"tags":1582,"stars":1548,"repoUrl":1549,"updatedAt":1593},"claude-api","build apps with the Claude API","Reference for the Claude API \u002F Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration.\nTRIGGER — read BEFORE opening the target file; don't skip because it \"looks like a one-liner\" — whenever: the prompt names Claude\u002FAnthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`); the user asks about an LLM (pricing\u002Fmodel choice\u002Flimits\u002Fcaching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent\u002FMCP\u002Ftool-definition\u002Fmulti-agent\u002FRAG\u002FLLM-judge\u002Fcomputer-use; generate\u002Fsummarize\u002Fextract\u002Fclassify\u002Frewrite\u002Fconverse over NL; debugging refusals\u002Fcutoffs\u002Fstreaming\u002Ftool-calls\u002Ftokens).\nSKIP only when another provider is being worked on (overrides all triggers): OpenAI\u002FGPT\u002FGemini\u002FLlama\u002FMistral\u002FCohere\u002FOllama named in the query; OR `grep -rE 'openai|langchain_openai|google.generativeai|genai|mistralai|cohere|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file).",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1583,1584,1585,1588,1590],{"name":1445,"slug":1446,"type":16},{"name":9,"slug":8,"type":16},{"name":1586,"slug":1587,"type":16},"Anthropic SDK","anthropic-sdk",{"name":1589,"slug":1578,"type":16},"Claude API",{"name":1591,"slug":1592,"type":16},"LLM","llm","2026-07-28T05:36:08.213335",{"slug":1595,"name":1595,"fn":1596,"description":1597,"org":1598,"tags":1599,"stars":1548,"repoUrl":1549,"updatedAt":1604},"doc-coauthoring","co-author documentation and technical specs","Guide users through a structured workflow for co-authoring documentation. Use when user wants to write documentation, proposals, technical specs, decision docs, or similar structured content. This workflow helps users efficiently transfer context, refine content through iteration, and verify the doc works for readers. Trigger when user mentions writing docs, creating proposals, drafting specs, or similar documentation tasks.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1600,1601],{"name":1449,"slug":1450,"type":16},{"name":1602,"slug":1603,"type":16},"Technical Writing","technical-writing","2026-04-06T17:56:14.18897",{"slug":1606,"name":1606,"fn":1607,"description":1608,"org":1609,"tags":1610,"stars":1548,"repoUrl":1549,"updatedAt":1625},"docx","create and edit Word documents","Use this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1611,1614,1616,1619,1622],{"name":1612,"slug":1613,"type":16},"Documents","documents",{"name":1615,"slug":1606,"type":16},"DOCX",{"name":1617,"slug":1618,"type":16},"Office","office",{"name":1620,"slug":1621,"type":16},"Templates","templates",{"name":1623,"slug":1624,"type":16},"Word","word","2026-07-18T05:16:23.136271",{"slug":1627,"name":1627,"fn":1628,"description":1629,"org":1630,"tags":1631,"stars":1548,"repoUrl":1549,"updatedAt":1643},"frontend-design","design production-grade frontend interfaces","Guidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1632,1633,1636,1639,1642],{"name":1540,"slug":1541,"type":16},{"name":1634,"slug":1635,"type":16},"Frontend","frontend",{"name":1637,"slug":1638,"type":16},"React","react",{"name":1640,"slug":1641,"type":16},"Tailwind CSS","tailwind-css",{"name":1467,"slug":1468,"type":16},"2026-04-06T17:56:16.723469",{"slug":1645,"name":1645,"fn":1646,"description":1647,"org":1648,"tags":1649,"stars":1548,"repoUrl":1549,"updatedAt":1657},"internal-comms","write internal company communications","A set of resources to help me write all kinds of internal communications, using the formats that my company likes to use. Claude should use this skill whenever asked to write some sort of internal communications (status reports, leadership updates, 3P updates, company newsletters, FAQs, incident reports, project updates, etc.).",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1650,1653,1654],{"name":1651,"slug":1652,"type":16},"Communications","communications",{"name":1620,"slug":1621,"type":16},{"name":1655,"slug":1656,"type":16},"Writing","writing","2026-04-06T17:56:20.695522",{"slug":1659,"name":1659,"fn":1472,"description":1660,"org":1661,"tags":1662,"stars":1548,"repoUrl":1549,"updatedAt":1667},"mcp-builder","Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node\u002FTypeScript (MCP SDK).",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1663,1664,1665,1666],{"name":1445,"slug":1446,"type":16},{"name":1478,"slug":1479,"type":16},{"name":1591,"slug":1592,"type":16},{"name":1465,"slug":27,"type":16},"2026-04-06T17:56:10.357665",{"slug":1575,"name":1575,"fn":1669,"description":1670,"org":1671,"tags":1672,"stars":1548,"repoUrl":1549,"updatedAt":1675},"read edit and manipulate PDF files","Use this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text\u002Ftables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting\u002Fdecrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1673,1674],{"name":1612,"slug":1613,"type":16},{"name":1574,"slug":1575,"type":16},"2026-04-06T17:56:02.483316",{"slug":1677,"name":1677,"fn":1678,"description":1679,"org":1680,"tags":1681,"stars":1548,"repoUrl":1549,"updatedAt":1688},"pptx","create and edit PowerPoint presentations","Use this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions \"deck,\" \"slides,\" \"presentation,\" or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1682,1685],{"name":1683,"slug":1684,"type":16},"PowerPoint","powerpoint",{"name":1686,"slug":1687,"type":16},"Presentations","presentations","2026-07-18T05:16:24.1471",{"slug":1690,"name":1690,"fn":1691,"description":1692,"org":1693,"tags":1694,"stars":1548,"repoUrl":1549,"updatedAt":1704},"skill-creator","create and optimize agent skills","Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1695,1696,1697,1700,1703],{"name":1445,"slug":1446,"type":16},{"name":1449,"slug":1450,"type":16},{"name":1698,"slug":1699,"type":16},"Evals","evals",{"name":1701,"slug":1702,"type":16},"Performance","performance",{"name":1602,"slug":1603,"type":16},"2026-04-19T06:45:40.804",490]