| name | heal-skill |
| description | Repair misfiring AI skill. |
| user-invocable | true |
| allowed-tools | ["Bash","Read","Write","Edit","Glob","Grep"] |
heal-skill
Close the feedback loop on skill quality: when a skill misbehaves in a real
session, repair it now, while the failure is in front of you. This is the
retrospective complement to authoring a skill and to ums (which records
learnings) — it fixes a skill that already shipped and then misfired.
When this fires
- A skill triggered on the wrong input, or failed to trigger on the right one.
- A skill fired correctly but its body led the work astray (bad steps, stale
instructions, wrong assumptions).
- The user says "that skill confused you", "heal that skill", "the description
is too greedy", "that shouldn't have fired", or
/heal-skill.
Procedure
1. Identify the culprit
Determine which skill misfired. If it's not named, infer it from what just
happened — the slash command invoked, the description that matched, or the
behavior that went wrong. Confirm the skill name with the user if ambiguous.
Read its skills/<name>/SKILL.md in full before changing anything.
2. Diagnose the root cause
Classify the failure — the fix differs by type:
- False trigger (fired when it shouldn't): the
description is too broad,
or its trigger phrases overlap a more general request. Fix: tighten the
description, narrow the trigger phrases, add an explicit "don't fire when…"
note.
- Missed trigger (didn't fire when it should): the
description lacks the
user's phrasing. Fix: add the missing trigger phrases / synonyms.
- Ambiguous overlap (two skills compete for the same request): fix the
boundary in both descriptions so each owns its lane; cross-link them.
- Bad body (fired right, acted wrong): a step is wrong, stale, or
underspecified. Fix the offending step; verify any file/flag/command it names
still exists.
State the diagnosis in one or two sentences before proposing the fix.
3. Propose a minimal fix
Draft the smallest change that addresses the root cause — usually a few lines
of the frontmatter description or one body step. Show the user the exact diff
and the reasoning. Don't rewrite the whole skill; don't add unrelated polish.
4. Apply with approval, then validate
On approval, edit the SKILL.md. Then sanity-check it still parses and is
internally consistent: run scripts/validate-skills.py if the repo has it, and
re-read the trigger against the failure that prompted the heal. If this repo
ships skills via branch + PR, follow that flow rather than committing to main.
5. Record the lesson
If the misfire reflects a pattern worth remembering (a recurring over-trigger,
a phrasing the user favors), hand off to ums to persist it, so the next skill
is authored without the same flaw.
Relationship to other skills
skill-builder — the authoring counterpart; it writes a skill, this skill
repairs one after it ships and misfires.
agent-builder — the authoring counterpart for a .claude/agents/*.md
subagent. If a heavy skill's fan-out step misbehaves because the spawned
agent's prompt or tool-scoping is wrong (not the calling skill's body), fix
the agent file via agent-builder's conventions instead of editing the
skill.
link-skills — the proactive, whole-corpus cross-link audit; heal-skill
is the single-skill reactive repair, and fixing an ambiguous overlap here
means cross-linking the two skills (exactly what link-skills sweeps for).
consolidate-skills — when two skills have real bodies for the same
workflow, that's a merge, not a heal; hand it there.
ums / record-learnings — hand off to persist a recurring misfire as a
remembered lesson, so the next skill is authored without the same flaw.
What NOT to do
- Don't delete or disable a skill to "fix" a single misfire — repair the
trigger or body instead. If a skill is genuinely redundant with another (two
real bodies for one workflow), that's a merge, not a heal — hand it to
consolidate-skills.
- Don't broaden a description so far that it starts stealing other skills'
triggers; healing a missed-trigger shouldn't create a false-trigger.
- Don't make unrelated edits in the same pass — keep the heal minimal and
reviewable.
- Don't change behavior the user was happy with; only touch what misfired.
The idea for this skill (a retrospective "repair the skill that just
confused you" loop) is credited in CREDITS.md.