# The Council — the portable core

*A share-friendly distillation of the `/council` skill from the Hey Kaleb newsletter. This is the method plus a prompt you can run in any capable model. The full skill is a multi-file engine — three modes, rubric-anchored seats, a saved-record system, and the research rubrics I use with clients — and those rubrics aren't in here. The part that travels is below. — Kaleb*

## The one rule that defines it

**Findings and fixes are separate stages, with a human decision between them.**

A reviewer that diagnoses and repairs in the same breath trains you to skip the diagnosis. That's the atrophy machine. Finding the flaw is the hard part you can't do from inside your own frame — worth handing off. Fixing it is where your judgment lives — hand that over every time and it quietly goes away. So the method stops after the findings and asks how you want to work them.

## How it runs

1. **Decorrelated read.** Several seats review the work, each pinned to a different part of a rubric so their *attention* diverges — different lenses, not five personalities. (The rubric split is the mechanism; the "personas" are a thin layer on top, and probably not what's doing the work.)
2. **A blind second pass.** A referee grades the arguments with the seat names stripped, so it judges the reasoning, not the reputation of a seat.
3. **The audit — the part most "AI review" skips.** Every load-bearing finding is checked against the actual source before you see it. Models import facts that were never there, get arithmetic wrong, and dismiss true findings. Mark each CONFIRMED / CORRECTED / OVERTURNED, and *show the overturned ones* — that's the proof the audit did work. A council you don't audit has told you five wrong things and sounded equally sure about all fourteen.
4. **Findings only.** Diagnosis in priority order, each marked *anchored* (you can point at the line) or *normative* (an opinion about good practice). No rewrite yet.
5. **The fork.** Stop and ask: *give me the fixes* / *coach me through them* (one at a time, questions not answers) / *coach the top one, fix the rest*. The last is usually right — it puts the effort where the judgment is, and doesn't turn a typo into a seminar.
6. **Advisory, always.** "Rework" is a suggestion; "Ship" is a read, not permission. You own the call.

**The one exception to the fork:** consent, privacy/PII, legal, security, and harm-to-others findings are never coached. They're stated directly, with the fix written out, up front — because the cost of you not getting there falls on someone who isn't in the room.

## The portable prompt

Paste this into any capable model with your work. It runs the method in one pass — a decorrelated read, a self-audit, findings only, then the fork.

```text
You are running a review council on the WORK below. Find what I can't see
from inside my own frame — then hand the findings back so I keep the
judgment. Follow this exactly.

STAKES: [one line — what this decides, what being wrong costs]

WORK:
[paste the plan, guide, draft, analysis, or decision here]

1. READ IT FROM FOUR ANGLES, committing fully to each before the next:
     - Correctness / risk: where does it break, what fragile assumption is
       load-bearing?
     - Framing: is this even the right problem, or a good answer to the
       wrong one?
     - Coverage: what's missing, unclaimed, or assumed?
     - The people affected: consent, privacy, who's absent from the sample,
       who bears the cost of being wrong?
   Ground every point in the actual work. Generic advice that would fit
   anything is the failure mode — cut it.

2. AUDIT YOURSELF before showing me anything. For each load-bearing point
   (one whose removal would change your verdict), go back to the WORK and
   check it: did you import a fact that isn't there? Is every number right?
   Mark each CONFIRMED, CORRECTED (give the correction), or OVERTURNED (cut
   it — but list what you cut and why). Don't quietly drop your own misses.

3. FINDINGS ONLY — do NOT rewrite the work. Return:
     - READ: Ship / Revise / Rework / Stop, one line with the reason. Any
       consent, privacy, legal, security, or harm blocker means it can't be
       "Ship."
     - MUST FIX FIRST: consent / privacy / legal / security / harm findings
       ONLY, each with the concrete fix written out (the actual clause, the
       field to cut). Skip this section if there are none.
     - WHAT BROKE: each finding in priority order — the problem, why it
       matters here, its audit mark, and ANCHORED (point at the line) or
       NORMATIVE (an opinion about good practice). No fixes yet.
     - WHAT HOLDS UP: briefly, what survived.

4. THEN STOP AND ASK how I want to work the findings:
     - Give me the fixes.
     - Coach me through them — one at a time, ask don't tell.
     - Coach the top one (the finding with the most judgment in it), fix the
       rest.
   Wait for my answer. Don't fix and ask in the same message.

Stay adversarial, but don't invent problems — if it's genuinely sound, say
so briefly and stop. Everything you return is advisory; I own the call.
```

## What this leaves out

The full skill routes by artifact type into three modes (council / panel / gate), carries rubric-anchored seat packs, ships a ready-to-run consent + AI-disclosure gate, and saves a retrievable record of every run. The research rubrics are how I work with teams — if that's what you need, [let's talk](https://www.heykaleb.com/aixuxr-transformation-chat).

## Credits & sources

The blind-review-plus-synthesis lineage traces to **Andrej Karpathy's "LLM Council"** ([github.com/karpathy/llm-council](https://github.com/karpathy/llm-council)) and the multi-agent-debate research it echoes — Du et al., *["Improving Factuality and Reasoning in Language Models through Multiagent Debate"](https://arxiv.org/abs/2305.14325)* (2023; ICML 2024) — plus the wider "LLM-as-a-judge" line. The through-line: independent critiques catch what a single, agreeable pass misses.

What's different here: Karpathy's council runs multiple *models* to *answer a question*. This runs decorrelated seats to *pressure-test your own work* — and adds the parts that matter for research integrity: the **findings/fixes fork** (so it amplifies your judgment instead of replacing it), the **source audit** (so it can't hand you confident, unverified findings), and **advisory-only** output. The lineage is theirs; the fork, the audit, and the honesty rules are mine.

Use it, adapt it, and pass the credit along.