| name | ai-harness-retrospective |
| description | Evaluate a real AI-assisted work session against the repo-local harness, classify reusable misses, and update or propose updates to docs, skills, scripts, MCP/tooling, or sensors. Use after substantial work, review feedback, failed validation, or human correction in a repo with an AI harness. |
AI Harness Retrospective
Use this after real work to keep a repo-local AI harness improving. The goal is
to convert recurring friction into deterministic harness changes without
turning one-off context into permanent process.
Read First
If present, read:
AGENTS.md
docs/ai-harness/harness-manifest.md
docs/ai-harness/README.md
- the skill or harness doc that should have guided the session
- the relevant diff, review comments, failed commands, or human corrections
If the repo has no harness skeleton, recommend ai-harness-bootstrap instead of
inventing an ad hoc retrospective structure.
Reconstruct The Session
Capture only the facts needed to evaluate the harness:
- task type and owning layer;
- files changed;
- skill or doc consulted;
- commands selected and run;
- review findings;
- human corrections or clarifications;
- generated-file, test-selection, source-boundary, or tooling mistakes;
- any repeated manual step that could be deterministic.
Classify Misses
For each meaningful miss, choose one bucket:
- doc update: missing or stale repo knowledge.
- skill update: wrong trigger, missing workflow step, or bad output shape.
- deterministic script/check: repeated mechanical step or drift check.
- MCP/tooling addition: external system or structured data access gap.
- sensor/evaluator: review or validation should catch this automatically.
- no harness change: one-off context, already covered guidance, or not worth
encoding.
Be conservative. Encode lessons that are likely to recur.
Update Rules
- Prefer patching an existing doc or skill before adding a new surface.
- Keep
AGENTS.md as a routing map, not a detailed playbook.
- Delete or shrink harness content that is generic, duplicated, or speculative.
- Update the harness manifest when active docs, skills, tools, generated outputs,
or known gaps change, if the repo has a manifest.
- If changing review automation or generated workflow sources, run the repo's
required compile/generation command.
- If the retrospective reveals a false harness claim, fix the claim before
adding new guidance.
- If a larger tool or sensor is needed, create issue text or acceptance criteria
instead of half-building it.
- If the only proposed guidance is "test your work", "read the docs", or
similarly generic advice, make no harness change.
Output
Return one of:
No harness change needed, with the reason.
- Exact doc or skill patch, with validation.
- Proposed script/tool/sensor, with acceptance criteria.
- Follow-up issue text for a larger harness improvement.
When edits are made, include:
- what session miss the edit addresses;
- why it is likely to recur;
- files changed;
- validation run;
- residual gaps.