Improve Skill

One cycle, one change, one independent verdict. You do not decide the skill got better.

Role

You are the orchestrator. You:

1. Resolve the target SKILL.md.
2. Snapshot the baseline off-tree before any edit.
3. Diagnose against the lens below and pick exactly one hole.
4. Apply that one change, or dispatch an author who does only that.
5. Dispatch a different agent as judge. Enforce the verdict.

You do not: score your own (or the author's) change; skip the gate; rewrite the skill; hand-edit generated README catalog files; create a skill from scratch.

NO INDEPENDENT PASS → NO COMMIT. FAIL → RESTORE BASELINE.

If you cannot dispatch a different agent, stop. Leave the tree at baseline. Do not self-score.

Inputs

Input	Required	Default
Target skill	no	The SKILL.md next to this file (improve-skill)
Commit prefix	no	none; the dimension still goes in the subject

If the user names a path that is not an existing */SKILL.md, stop and say so. Do not ask them what to change.

Out of scope: README, cheatsheet, harness files, new skills, application code, duration-boxed work.

Working files

$d is created in step 2 (/tmp/improve-skill-<target-basename>-<epoch>/).

baseline/          # committed (or pre-edit) copy of the skill tree
diagnosis.md       # one hole, or "nothing above the bar"
gate.md            # judge verdict — written by the judge, not you
log.md             # append-only

Pass paths, not file contents, to the author and the judge.

references/ next to this file are procedure helpers. $skill_dir/references/ is the target's. They coincide only when the target is this skill.

Process

select-target → baseline → diagnose → one-change → independent-gate
  PASS: regen catalog if needed, commit, record hash
  FAIL: restore baseline, do not commit, record why
  nothing-to-do: stop, no commit (valid refused improvement)

1. Select target

Start-here checklist — first match wins:

1. If the user named more than one skill (a list, "all skills", more than one SKILL.md or skill directory), stop. One cycle accepts one target. Do not pick a favorite. Do not ask which.
2. If the user named a path that is not an existing */SKILL.md and not a skill directory that contains one, stop and say so. Do not default to this skill. Do not ask them what to change.
3. If the user gave an existing …/SKILL.md (or a skill directory that contains one), that is the target.
4. Otherwise the target is this skill: the SKILL.md in the same directory as the file you are reading.
5. Confirm the file exists and YAML name: equals the directory name. If not, stop.

2. Capture baseline

Do this before any edit. skill_dir is the absolute path to the directory that contains the target SKILL.md. Run from a shell:

skill_dir=<absolute path>
slug=$(basename "$skill_dir")
d=/tmp/improve-skill-$slug-$(date +%s)
mkdir -p "$d/baseline"
repo=$(git -C "$skill_dir" rev-parse --show-toplevel)
rel=$(python3 -c "import os; print(os.path.relpath('$skill_dir', '$repo'))")
if git -C "$repo" rev-parse --verify --quiet "HEAD:$rel" >/dev/null; then
  git -C "$repo" archive HEAD "$rel" | tar -x -C "$d/baseline" --strip-components=1
else
  cp -R "$skill_dir/." "$d/baseline/"
fi
git -C "$repo" rev-parse --short HEAD > "$d/baseline-rev.txt"

If HEAD:$rel exists, that tree is the baseline even when the working tree is dirty. If it does not, the live $skill_dir tree is the baseline — not only SKILL.md + references/. Restore uses $d/baseline/. If $d/baseline/SKILL.md is missing, stop. Do not edit. Do not invent a baseline.

3. Diagnose

Read the target SKILL.md and any references/ one level under it. Write $d/diagnosis.md. Do not ask the user what is wrong.

Lens (apply all; pick one hole):

1. Description = WHAT + WHEN. Triggers named. Siblings that could steal load are excluded by name.
2. Executable procedure. A cold agent can name the next action and the stop/fail condition. Ordered steps, not principles.
3. Degrees of freedom. Fragile steps (baseline, one-change, independent gate, commit-on-pass) are locked. Wording of a change is free. "Use judgment" on a gated step is a hole.
4. One home per fact; no sprawl. No tutorial on what a skill is. No no-op restatement.
5. Four rubric dimensions (the judge scores these same four):
   • Clarity — every step names a next action or a stop condition.
   • Trigger reliability — intended utterances land here; sibling collisions do not.
   • Procedure completeness — no step the skill claims to require is missing or "figure it out".
   • Anti-failure coverage — each named failure mode has a named protection (restore, refuse, stop).

A hole is above the bar only if you can quote a missing next action, stop/fail condition, trigger or sibling exclusion, or named protection. Rephrase, shorter wording, synonym, extra example of an existing rule, restating a protection already in the table, or a user ask that is only polish / length / wording cleanup is not a hole — write nothing above the bar.

Write exactly one of:

# Diagnosis
- Dimension or failure mode: <one name>
- Hole: "<quote from the current skill or ABSENT>"
- Change: <one sentence>

or

# Diagnosis
nothing above the bar

If nothing above the bar, append that to log.md and stop. Do not edit. That is a valid cycle.

4. One focused change

Edit only the hole named in diagnosis.md. One dimension or one failure mode. Not a rewrite. Not "also cleaned up X".

• You may edit SKILL.md yourself or dispatch an author with the target path, the diagnosis path, and "change only this hole".
• Do not edit generated README.md files.
• After the edit, if $repo/scripts/generate.ts exists, run from $repo:

deno run --allow-read --allow-write scripts/generate.ts

Never hand-patch the root ### Skills list.

Then revert extras before gating. Reuse $skill_dir, $d, and $repo from step 2. If $skill_dir is unset or "$skill_dir/SKILL.md" is not the live candidate, stop. Do not invent a revert.

An extra is a dirty path under $skill_dir that is not $skill_dir/SKILL.md, not $skill_dir/README.md, and not backtick-quoted (relative to $skill_dir, or as $skill_dir/...) in the - Change: line of $d/diagnosis.md. Generated $repo/README.md is not an extra. A new or non-SKILL.md path is kept only when that Change line quotes it.

change=$(sed -n 's/^- Change: //p' "$d/diagnosis.md")
prefix=$(python3 -c "import os; print(os.path.relpath('$skill_dir', '$repo'))")
git -C "$repo" status --porcelain -u -- "$skill_dir" | while IFS= read -r line; do
  path="${line#???}"
  rel="${path#"$prefix"/}"
  if [ "$rel" = "SKILL.md" ] || [ "$rel" = "README.md" ]; then
    continue
  fi
  case "$change" in
    *'`'"$rel"'`'*) continue ;;
    *'`'"$skill_dir/$rel"'`'*) continue ;;
  esac
  if [ -e "$d/baseline/$rel" ]; then
    git -C "$repo" checkout HEAD -- "$path"
  else
    rm -rf "$repo/$path"
  fi
done

5. Independent gate

Dispatch a new agent. Not you. Not the author. Give it only the baseline path, the live candidate path, this section, and the absolute path of references/known-bad.md next to this file.

Judge prompt (fill the brackets, keep the rest):

You are an independent skill judge. You did not author this change.
Baseline: < $d/baseline/SKILL.md >
Candidate: < live SKILL.md path >
Known-bad fixtures: <absolute path of references/known-bad.md next to this skill>
Also read sibling files in each tree if present.

Do not improve the candidate. Do not take the author's word.

1. Mechanical validity — FAIL if any miss:
   - candidate exists; directory name == YAML name
   - YAML name: and description: are single-line (not `>` / `|`)
   - description is 1–1024 chars, no XML tags, states WHAT and WHEN
   - a generate.ts-style first-line parse would capture the real description text
2. Score each dimension BETTER / SAME / WORSE. Quote both versions
   (or "absent in baseline") for every cell.
   - Clarity: next action or stop condition newly named vs buried or vaguer
   - Trigger reliability: intended utterance newly caught, or sibling newly
     excluded vs lost trigger or new collision
   - Procedure completeness: a required step newly specified vs removed or
     weakened to judgment
   - Anti-failure coverage: a named failure mode newly protected vs a
     protection removed
3. Known-bad — read that fixtures file. FAIL if the candidate matches any
   fixture (same shape, not only byte-identical). Do not invent shapes.
   Do not skip a fixture because the candidate added sentences.

Write < $d/gate.md > as:
- validity: PASS or FAIL + first miss
- four cells with quotes
- known-bad: match/no-match per shape
- Verdict: PASS or FAIL

PASS only if validity holds AND at least one BETTER AND zero WORSE AND no
known-bad match. 0 BETTER (rephrase, length-only, synonym churn) is FAIL.
Length is not a dimension. "Looks better" without quotes is FAIL.

You do not write gate.md. After the judge returns, go to step 6 even when $d/gate.md or a Verdict: line is missing. Do not invent either.

6. Commit only on PASS

Reuse $skill_dir, $d, and $repo from step 2. If $skill_dir is unset or "$skill_dir/SKILL.md" is not the live candidate, stop. Do not invent an add or restore path.

Parse $d/gate.md if it exists. Do not write it. PASS only when a Verdict: PASS line is present and no Verdict: FAIL line is present. Otherwise FAIL — missing file, missing Verdict:, unparseable, or FAIL.

PASS

Add only $skill_dir/SKILL.md, $skill_dir/README.md if present, paths backtick-quoted in the - Change: line (same contract as step 4), and $repo/README.md if generate.ts exists. Do not git add "$skill_dir". If any other dirty path remains under $skill_dir, stop — do not commit.

change=$(sed -n 's/^- Change: //p' "$d/diagnosis.md")
prefix=$(python3 -c "import os; print(os.path.relpath('$skill_dir', '$repo'))")
git -C "$repo" add -- "$skill_dir/SKILL.md"
if [ -f "$skill_dir/README.md" ]; then git -C "$repo" add -- "$skill_dir/README.md"; fi
if [ -f "$repo/scripts/generate.ts" ]; then git -C "$repo" add -- README.md; fi
git -C "$repo" status --porcelain -u -- "$skill_dir" | while IFS= read -r line; do
  path="${line#???}"
  rel="${path#"$prefix"/}"
  case "$change" in
    *'`'"$rel"'`'*) git -C "$repo" add -- "$path" ;;
    *'`'"$skill_dir/$rel"'`'*) git -C "$repo" add -- "$path" ;;
  esac
done
if git -C "$repo" status --porcelain -u -- "$skill_dir" | grep -q '^.[^ ]'; then
  echo "extras remain under $skill_dir; stop; do not commit"
  exit 1
fi
git -C "$repo" commit -m "<prefix>improve-skill: <dimension> — <one line>"

<dimension> is the name from diagnosis.md. <prefix> is the caller's required prefix plus a trailing space, or empty.

Record the hash in log.md.

FAIL

Run the FAIL restore in references/restore.md next to this file (not $skill_dir): overlay $d/baseline/, then delete live paths with no baseline counterpart. Stop if $d/baseline/SKILL.md is missing. Then:

if [ -f "$repo/scripts/generate.ts" ]; then
  (cd "$repo" && deno run --allow-read --allow-write scripts/generate.ts)
fi

Do not commit. Record the judge's first failing clause in log.md, or missing Verdict: if the file or Verdict: line is absent.

You are not the judge. Do not override FAIL because the diff looks fine.

Failure modes

Failure mode	Protection
Author declares the new version better	Judge is a different agent; self-score is not a verdict
No subagent available	Stop at baseline; do not self-score
Kitchen-sink rewrite	Diagnosis names one hole; step 4 reverts dirty $skill_dir paths that are not SKILL.md, not generated README.md, and not Change-quoted
PASS add stages extras	Do not git add "$skill_dir"; add SKILL.md, generated README.md, and Change-quoted paths only; stop if extras remain
Edit with no snapshot	Step 2 runs before any write; stop if $d/baseline/SKILL.md is missing; restore path is $d/baseline/
Commit then maybe measure	Step 6 reads gate.md first; FAIL does not git commit
Judge returns no gate.md / no Verdict:	Step 6 still runs; missing or unparseable → FAIL restore; do not write gate.md; do not commit
FAIL restore or PASS add hits the wrong tree	Step 6 reuses $skill_dir from step 2; stop if unset or not the live candidate
FAIL restore leaves newly added references/	Always rm -rf "$skill_dir/references"; copy from $d/baseline/references only if that dir exists
FAIL restore leaves newly added files outside references/	Run references/restore.md next to this file: overlay $d/baseline/, delete live paths with no baseline counterpart
Untracked fallback omits assets/ or extra root files	Step 2 else copies the whole $skill_dir tree; FAIL extra-delete cannot drop pre-existing extras
Catalog description becomes >	Single-line description: required; generate.ts after description edits
Hand-edited README catalog	generate.ts only; never patch ### Skills by hand
Ask the user what to fix	Default target; lens produces the hole or a refusal
Improve a non-skill	Step 1 stops unless the path is an existing SKILL.md
More than one target	Step 1 stops; one cycle accepts one target
Sibling collision (duration work, code slop, graphs, metric loops, prose deslop)	Description excludes anneal, timeboxed-iterating, workgraph, autoresearch, catalog-deslop
Known-bad scored from memory	Judge reads this skill's references/known-bad.md; any fixture match → FAIL
Cosmetic-only / rephrase cycle	Step 3 refuses unless the hole quotes a missing action, stop, exclusion, or protection; user-asked polish is nothing above the bar

Red flags — stop and reread

• You are about to commit without a Verdict: PASS written by someone else.
• $d/gate.md is missing (or has no Verdict:) and you are about to commit or leave the candidate in the tree.
• You scored the change yourself.
• diagnosis.md lists more than one hole and you are fixing them all.
• You asked the user what to change.
• You are hand-editing a README.
• The working tree still has the candidate after a FAIL.
• You skipped baseline because it is a small edit.
• $d/baseline/SKILL.md is missing and you are about to edit.
• The target is untracked and $d/baseline/ is missing live assets/ (or extra root files) and you are about to edit.
• $skill_dir is unset (or not the live candidate) and you are about to git add or restore.
• Live references/ is still present after a FAIL and $d/baseline/references was missing.
• Live assets/ or an extra root file is still present after a FAIL and $d/baseline/ did not have it.
• You are scoring known-bad from memory instead of reading this skill's references/known-bad.md.
• The hole is a wording preference (rephrase, shorter, synonym, restated protection) and you are about to edit anyway.
• The user named two skills (or all skills) and you are about to pick one or run both.
• The user named a path that is not an existing SKILL.md (or skill directory) and you are about to default to this skill.
• Dirty extras remain under $skill_dir (not SKILL.md, not generated README.md, not Change-quoted) and you are about to gate.
• You are about to git add "$skill_dir" (the whole directory) or extras remain under $skill_dir and you are about to commit.