Rick Improve

Rick Improve

Bounded self-improvement loop on a Rick skill prompt. The prompt is the source code — this is how you refactor it without losing the file to drift.

Target skill: $ARGUMENTS

If $ARGUMENTS is empty, tell Morty to give a target skill name and stop. Don't guess.

Resolve the target

Look in priority order, use the first match:

1. An absolute path to the skills source repo if Morty's memory, CLAUDE.md, or this conversation references one (e.g. ~/code/deren-skills/skills/rick/$ARGUMENTS/SKILL.md). Installed copies get clobbered on resync — always prefer the source of truth when known.
2. skills/rick/$ARGUMENTS/SKILL.md (source repo, relative to cwd)
3. .claude/skills/$ARGUMENTS/SKILL.md (installed location, last resort)

If none exists, list available rick skill directories from whichever root resolved and stop. Don't try to be clever.

Below, <target-dir> is the directory holding the resolved SKILL.md.

Reference bar

Read rick-mode/SKILL.md from the same root as the resolved target (i.e. the rick-mode/ sibling of <target-dir>). It is the gold standard: unambiguous enforcement, no vague bans, examples where format matters, gates with teeth. Use it as the bar for every round's critique. Anything the target does worse than rick-mode is a candidate finding.

Before the loop

1. Read the target SKILL.md. The actual file.
2. Determine version number: list <target-dir>/versions/v*.md, find the highest integer (ignore -pre suffix), next = N+1. If the directory or files don't exist, this is v1.
3. Snapshot: copy current SKILL.md → <target-dir>/versions/v<N>-pre.md. Create the versions/ dir if needed.
4. Append a stub line to <target-dir>/versions/VERSIONS.md (create with the header below if missing): - v<N> | <YYYY-MM-DD> | pre-loop snapshot
5. Count body lines (excluding frontmatter). If already over 250, the first round must trim before it can add anything.

Stop conditions (first one that fires, stop immediately)

1. 3 rounds completed. Hard cap.
2. Round produces zero net-new changes. Trivial polish (whitespace, word swaps without semantic change) does not count. Say "no improvements found this round" and stop.
3. Body exceeds 250 lines. Skill files are read on-demand, but this is the ceiling where the model starts ignoring sections.
4. A finding needs a whole-section rewrite. That's a different skill (/rick-plan-improve), not this one. Still snapshot post (so the partial round's edits are captured) and tag the entry needs-rewrite.

Loop protocol (repeat up to 3 times)

3 is a cap, not a target. Most skills converge in 1-2 rounds. If round 2 or 3 starts and you cannot name a real failure mode, stop early — don't dig for micro-polish to "use the budget."

Each round:

1. Read the current file. The actual file, not your memory of it.

2. Critique against the weakness categories below. Every finding uses this format:

   [weakness category from list below]

   • Where: SKILL.md:<line> or quoted snippet
   • Failure mode: one sentence naming what the model does wrong because of this

   Example:

   Vague enforcement

   • Where: SKILL.md:42 — "be specific in your critique"
   • Failure mode: model writes "this section is unclear" with no line citation, leaving the next round nothing concrete to act on.

3. Rank problems worst first. If you can't fill the Failure mode line, the problem does not qualify. Skip it.

4. Apply surgical fixes. Edit specific lines, not whole sections. (Whole-section rewrite triggers Stop condition 4.)

5. Count body lines after edits. If over 250, stop the loop immediately regardless of round number.

6. Report what changed and why. One line per change. Hold these for the post-loop VERSIONS.md entry.

After the loop

1. Snapshot: copy final SKILL.md → <target-dir>/versions/v<N>.md (post-loop result, bare integer matches the convention used by rick-plan-improve and rick-review).
2. Append the full entry to <target-dir>/versions/VERSIONS.md using the format below.

Weakness categories

• Redundancy: two sections saying the same thing. The model picks one and ignores the other.
• Missing examples: a format requirement with no example. The model freestyles it wrong.
• Unconditional rules that should be conditional: "always do X" when X should only fire in specific situations.
• Buried critical instructions: anything that must survive context compression needs its own labeled block, not buried in a paragraph.
• Vague enforcement: "be specific" is not enforceable. "cite file:line or paste the snippet" is.
• Overlapping sections: if two sections could swap content without anyone noticing, differentiate them or merge.
• Rules with no teeth: a rule that says what to do but not what happens if you don't. Add the consequence.
• Manufacturing pressure: phrasing that implies "keep looking until you find something." Creates pressure to invent findings.

What NOT to do

• Don't add sections because they seem useful. Every section must prevent a specific failure mode you can name.
• Don't expand examples into tutorials. One tight example per format requirement.
• Don't touch versions/ files mid-loop. Only write pre (before round 1) and post (after the final round).
• Don't change the skill's frontmatter name field. Renaming the skill breaks every downstream reference.

VERSIONS.md format

See VERSIONS-FORMAT.md for the file header, pre-loop stub shape, post-loop full entry shape, and restore-condition rules. Read it before the first write (the pre-loop stub in step 4 of "Before the loop").

Confirm and stop

After the post snapshot is written and VERSIONS.md is updated, print exactly:

Improved <skill-name> (v<N>): <stop-condition>. History: <target-dir>/versions/VERSIONS.md

Then stop. The round-by-round summary is already in VERSIONS.md. No recap. No "want me to revert."