docs(ce-skill-work): size a behavioral eval to the decision, not the workflow

[tmchow]

ce-skill-work's evaluate reference now says how big a behavioral eval of a routing or activation change should be: stop the cell at the decision it proves and grade the declared line, one discriminating fixture per direction (plus one per extra leg of the condition), post arm on Claude and Codex, pre arm only when no baseline evidence exists, six to twelve concurrent cells. Until now the guide said "proportionally to risk" with no default, and the eval on #1517 defaulted to full ce-doc-review runs per cell — ten minutes each, serial, across a fixture matrix — to observe a team announcement decided in the first thirty seconds.

Repo-local authoring guidance only; no shipped skill changes. tests/repo-local-ce-skill-work.test.ts passes.

Security Disclosure

No security-relevant changes.

Agent Disclosure

• Model: Claude Code · claude-fable-5

---

[cursor]

PR Summary

Low Risk
Documentation-only change to authoring eval sizing; no shipped skill, runtime, or security behavior.

Overview
Tells skill authors to stop a routing/activation eval at the decision it proves, instead of running a full skill workflow per cell.

The new default: one discriminating fixture per direction (plus one per extra condition leg), grade the declared result not narration, post-arm on Claude and Codex, pre-arm only when no session baseline exists, and keep packs around six to twelve cells. Activation still needs a fresh host session with no skill injected. Repo-local authoring guidance only.

^{Reviewed by Cursor Bugbot for commit 339915d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aa117ed2b4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".