Develop behavior changes by first observing an expected test failure, then making the smallest passing implementation. Refactor with tests passing, and document preserved behavior and measured runtime whenever tests are consolidated.
Build a behavior change with observed red, minimal green, and measured test consolidation
Review the instructions the agent will receive before you install.
Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than
/octo:*slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, seeskills/blocks/codex-host-adapter.md.
Read skills/blocks/engineering-method-selection.md from the installed plugin
for review admission. Natural-language requests and --peer-review share that
policy. Honor host-only requests; risk alone does not authorize paid usage.
Run the red, green, and refactor cycle on the current host. Routine TDD makes
zero additional provider dispatches. Use one external reviewer only when the
user passes --peer-review, explicitly requests independent review, or an
existing risk policy requires it. Explicit debate, council, and multi-model
commands retain their own execution contracts.
Do not change production behavior until a focused test fails for the expected reason. Existing implementation outside the requested change remains intact.
If a test passes before implementation, it is not red evidence. If it errors due to fixture or syntax problems, repair the test until it fails on the missing behavior.
Do not equate similar assertions with duplicate guarantees. Keep separate OS,
security, cancellation, and integration boundaries. For every removed test,
record old_test, behavior, replacement, mutant, red_observed,
baseline_ms, candidate_ms, and reason. The retained test must kill the
named mutant at the intended caller boundary.
After one warm-up, measure five isolated runs and report every sample and the median. Review a slowdown only when it exceeds both 20 percent and 100 ms.
Completion requires observed red and green evidence, affected-suite results, and the consolidation ledger when tests were removed. A missing reviewer is reported as incomplete review, never simulated.
If the same test remains red after two implementation attempts, stop and recheck the test boundary, fixture, and expected behavior. The strategy-rotation hook is a signal to try a fundamentally different hypothesis, not another variation of the same patch.
Adapted from DEEPENING in mattpocock/skills at commit
3cca18b368ae95cdbdebbff572ccafa662551015 under the MIT License. See
THIRD_PARTY_NOTICES.md.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Interpret the saved results of a completed AI council review. Count verdicts, check whether enough participants responded and whether models are independent, surface substantive dissent, and explain what action the outcome supports.
Turn a technical choice into a structured debate between different AI model families. Assign explicit positions, launch the discussion, and summarize agreement, disagreement, and a recommendation supported by the strongest argument that survives the debate.
Clarify a feature’s requirements with multiple AI providers, combining technical constraints with user and business needs. Produce a scoped definition covering priorities, edge cases, exclusions, and a requirements checklist before implementation begins.
Review code or documents with available AI providers before delivery. Combine findings into a scored validation report that checks prior requirements and decisions, separates critical issues from warnings, and recommends next steps.
Develop complex features or business documents by gathering approaches from available AI providers. Combine their perspectives with prior project decisions and quality checks to produce a proposed plan, then proceed with implementation after user confirmation.
Research technical or business questions through multiple available AI providers, each assigned a different perspective. Combine usable results into a report covering findings, disagreements, evidence gaps, priorities, and recommended next steps.