Skill Refine
Convert feedback into plan-level patches. Do not edit the skill bundle directly.
Quick start
Agent: "refine plan skills/my-skill/.skill-plan.yaml"
Workflows
1. Load plan + reports
- Require existing
.skill-plan.yamlpath. If missing, stop and ask user for path. - Read base plan YAML. Confirm it is parseable and has required planning fields.
- Locate
*-skill-check-report.htmland/or*-eval-report.htmlprovided by user. - Run
Bash scripts/parse-report.sh <report-path>for each report. - Validate parsed JSON against
assets/report-schema.json. Show parse errors if malformed. - Present feedback summary: total findings, top failures, affected plan areas.
2. Rank findings
- Read
references/priority-rules.mdand classify findings into P0-P3. - Sort findings by tier, then impact within tier.
- Present ranked queue to user. Include counts per priority tier.
3. Propose plan patch operations
- For top-ranked finding, generate one plan patch proposal using
references/change-proposal-format.md. - Scope patches to plan fields only (e.g.,
intent,scope,mechanism,context_assets,workflow_notes,example_placed,target_path). - Present options:
[apply / skip / edit / stop]. - On
apply, append operation to patch queue. - On
edit, accept user's modified patch operation and re-present once for confirmation. - On
skip, advance to next finding. Onstop, persist queue and exit loop.
4. Build artifacts
- Write queued operations to
<plan-dir>/.skill-plan.patch.yamlusingassets/plan-patch-template.yamlshape. - Apply operations to base plan in memory and produce
<plan-dir>/.skill-plan.patched.yaml. - Run
Bash scripts/validate-plan.sh <plan-dir>/.skill-plan.patched.yamlto validate forge import contract. - If validation fails, mark failing operations and ask user whether to edit/drop them.
- On validation pass, show concise plan diff summary (changed fields only).
5. Handoff
- Ask user to approve patched plan for forge input.
- On approval, instruct:
/skill-forge <plan-dir>/.skill-plan.patched.yaml. - On reject, keep patch artifact and return to step 10.
Guardrails
- Human-in-the-loop always. Never auto-apply patches without explicit user approval.
- Never edit target skill bundle files directly in this skill.
- Keep patch operations granular (one concern per operation).
- Preserve original
.skill-plan.yaml; patched output goes to.skill-plan.patched.yaml.
Reference files
See priority-rules.md for finding severity tiers and ranking logic. See change-proposal-format.md for patch proposal template. See Example.md for end-to-end walkthrough. See report-schema.json for parsed report JSON schema. See plan-patch-template.yaml for patch artifact shape. See changelog-template.md for patch log entry format.