agentleFS
naman10parikhnaman10parikh/penny/.claude/skills/llm-council-critique/SKILL.md

llm-council-critique

Skill0 starsChanged 28 days agoMITView on GitHub

Run a 6-hat brutal LLM Council self-critique on any deliverable, session output, or status update. Surfaces what's actually missing, where the work is shallow, and what was tabled-for-later. Use BEFORE re-attacking a problem the chairman has flagged twice. Encodes the Karpathy LLM Council pattern from CP82 ralph loop iterations 2 and 3 — applying the council BEFORE the next attack saves three iterations.

4 other projects use a version of this, 2 of them edited.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before any "I think we're done" claim on a complex deliverable

    line 15
  2. Distinctive:

    After spawning sub-agents and getting output back — BEFORE applying their proposals

    line 16

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

LLM Council Critique — Brutal Self-Reflection Skill

Permanent harness asset for the 6-hat brutal self-critique pattern Karpathy popularized and chairman invoked in CP82 ralph loop iterations 2 and 3.

When to invoke

  • Chairman says "be very blunt with yourself, I'm not seeing it"
  • Chairman re-invokes the same prompt for a 2nd or 3rd time (proves prior iteration was insufficient)
  • Before any "I think we're done" claim on a complex deliverable
  • After spawning sub-agents and getting output back — BEFORE applying their proposals
  • Anytime you catch yourself listing "follow-ups" or "still ungated" items at the end of a deliverable

The 6 hats (always run in this order)

Hat 1: Senior FDE / Architect

Lens: did you ship a SYSTEM or just words? What's the gap between "I added prose" and "the platform is buildable from this"? Are there scenes/sections/files you avoided?

Hat 2: Boris Cherny / Harness Engineer

Lens: did you DOGFOOD the framework you described? If you added a "5-axis Observer" claim to the doc, does Energy's actual harness implement it? Where's the test that proves the prose isn't aspirational?

Hat 3: Karpathy / Open-Endedness

Lens: did you USE the framework as a generative lens or just CITE it? Open-endedness means: when you ask "what's genuinely different from anything we already do?" — what does the answer surface? Did you generate any genuinely new ideas, or just stitch existing material?

Hat 4: Auditor / Coverage

Lens: count what was asked vs done. If chairman said "every bookmark" and you processed 30 of 117 (25%), that's a 75% miss. If chairman said "create tools out of every single thing I've told you" and you created ZERO, that's a hard miss. Be quantitative.

Hat 5: Process Skeptic

Lens: is iteration N shaped the SAME as iteration N-1? If yes, you're loop-aware but loop-trapped. To make progress, attack a NEW SURFACE (different sections, different files, different artifact types). Same-shape iteration with fancier prose isn't progress.

Hat 6: Chairman's voice

Lens: read the literal words of the chairman's prompt. What did they ask for that you SKIPPED? Words like "every", "all", "create tools out of every single thing", "boil the ocean", "don't stop until done" are LITERAL — not aspirational. Where did you treat a literal directive as aspirational?

Output format

## LLM Council Critique — Iteration N

**Hat 1 (FDE):** "<one-sentence brutal observation>"
**Hat 2 (Boris Cherny):** "<one-sentence brutal observation>"
**Hat 3 (Karpathy):** "<one-sentence brutal observation>"
**Hat 4 (Auditor):** "<one-sentence brutal observation with quantification>"
**Hat 5 (Process Skeptic):** "<one-sentence brutal observation>"
**Hat 6 (Chairman):** "<one-sentence brutal observation>"

## What's actually left (the new attack surface)

1. <specific deliverable, not category>
2. <specific deliverable>
3. ...

## Iteration N+1 plan (DIFFERENT shape from N)

<2-3 sentences on what makes this iteration structurally different>

Anti-patterns (banned)

  • ❌ "Soft" critiques ("this is good but could be better") — must be brutal, must name specific gaps
  • ❌ "Things to consider" or "follow-ups" — the WHOLE POINT is to surface what's left and DO it, not list it
  • ❌ Repeating an iteration N-1 critique in iteration N (means you didn't actually fix the prior issues)
  • ❌ Running the council AFTER claiming completion (run it BEFORE — saves 3 iterations)
  • ❌ Skipping any of the 6 hats (each surfaces a different blind spot; missing hats = missed gaps)

When NOT to invoke

  • Routine bug fixes with no chairman directive
  • Single-file edits where the scope is unambiguous
  • Tasks the chairman explicitly closed ("ship it", "good enough", "stop here")

Why this skill exists (provenance)

Chairman invoked the same Ralph Loop prompt twice in CP82 (iterations 2 and 3) because each time the prior iteration's "follow-up list" at the end was a tell that the work was incomplete. Iteration 3's attack surface (Scenes/Mechanics + new skills + Appendix + vault MOC) was structurally different from iter 1+2 (which both attacked TLDR sections). The skill codifies: when chairman re-invokes, run the council BEFORE the next attack — surface what was missed, then attack a NEW surface.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.