Write an agent behavior spec

Contract: references/interaction-contract.md. Calibration, templates, provenance: references/behavior-spec-template.md.

Trigger

• Requests naming BEHAVIOR.md, a behavioral contract, or recurring agent conduct.
• Review expectations that live only in reviewers' heads.
• A failure class recurring across prompts and models — the standard is missing, not the prompt.

Do

1. Name the recurring behavior and why mistakes matter. If traces exist, infer the spec from them before asking anything.
2. Write the seven sections in reference.md: intent, applicability, evidence, decision, execution, recovery, failure modes.
3. Write every clause so a reviewer could mark it satisfied or violated from a trace alone. Replace "handles uncertainty well" with "states the ambiguity and asks before writing files."
4. Keep clauses durable across a prompt rewrite or model swap. Anything that would not survive one belongs in the prompt, not the spec.
5. Version and date it, then emit the observable review questions the spec implies — the contract that evidence mapping and scorer work consume.

Avoid

• Do not turn the spec into a system prompt, tool manual, or task instruction; it outlives all three.
• Do not encode tool names, model strings, or prompt wording into durable clauses.
• Do not enumerate every situation; specify the decision rule and its boundaries.
• Do not let the agent's author write the failure-modes section.

Check

• All seven sections present; every clause observable in a trace and capable of being violated.
• Failure modes named and distinct, not a generic "does the wrong thing."
• Versioned, dated, paired with review questions for the next skill.

Risk

• Vague conduct produces subjective review, and subjective review produces disagreement that gets misread as model variance.
• Over-specification makes the standard brittle: a spec pinned to today's tool set fails the day a tool is renamed.
• An orphaned spec is the common failure — written, admired, never scored against. The review questions are what prevent it.

Braintrust

A spec is not a platform object; its clauses become three that are. Each review question becomes a per-criterion score in a human review queue with free-text notes — this is how the spec starts producing labels instead of sitting in a doc. Each failure mode becomes its own scorer (references/platform-mechanics.md §5), so a regression in one mode is its own column rather than averaged away. Applied to live traffic, violations become trace classifications you can alert and slice on. Tag the spec version in scorer span metadata, and treat a spec-version boundary like a scorer-version boundary: do not compare across it. When a review queue surfaces a violation the spec does not cover, that is a spec gap — route it back here rather than widening a scorer.