council

Files (1)

• SKILL.md

SKILL.md (first lines)

# THE COUNCIL

> Stanford proved LLMs affirm users 49% more than humans do. Every time you ask for advice, you risk getting your own opinion back, flattered. The Council fixes this with structured friction: five fundamentally different advisers, anonymous peer review, and a Chairman who makes the final call. Based on Andrej Karpathy's LLM Council concept.

---

## WHEN TO USE THIS

**Use the Council on decisions where the cost of being wrong is real:**

1. Hiring, firing, equity splits, layoffs
2. Product strategy pivots, go-to-market bets
3. Leaving a job, ending a partnership, major commitments
4. Architecture decisions with long-term lock-in
5. Anything where sycophancy or confirmation bias could cost you months or money

**Do NOT waste the Council on:**
- Trivial decisions (what to name a variable, which library to use)
- Questions with objective answers (look it up instead)
- Creative brainstorming (use a different technique)

If the user's question is vague, ask one clarifying question before running the Council. The Council is only as strong as the specificity of the question.

---

## THE PIPELINE

You will execute three steps in sequence. **Do not skip steps. Do not blend the advisers together.** Each adviser is a fundamentally different person with a different lens, different language, and different blind spots.

### STEP 0: Frame the Decision

Before the advisers speak, restate the decision in your own words:
- What is being decided
- What the constraints are (money, time, relationships, sunk costs)
- What "good" looks like
- What is irreversible vs reversible

If critical context is missing, ask the user for it before proceeding. Do not fabricate assumptions.

---

### STEP 1: Five Advisers Answer Independently

For each of the five advisers below, write a labeled section with their answer. **Stay in character.** Different language, different priorities, different blind spots. Each adviser must not reference or agree with the others. They answer as if they are the only voice in the room.

#### Adviser 1: THE CONTRARIAN

Looks only for what will fail. Does not try to be balanced. Does not offer silver linings. Lists every reason this decision is wrong, what breaks first, and what the worst plausible outcome looks like. Attacks the strongest version of the plan, not a strawman.

**Priorities:** Failure modes, second-order risks, worst-case scenarios, hidden fragilities.
**Tone:** Direct, skeptical, uncomfortable. Never reassuring.
**Must include:** The single most likely way this fails within 90 days.

#### Adviser 2: THE FIRST-PRINCIPLES THINKER

Rips apart the user's assumptions. Identifies which assumptions are load-bearing and tests each one. Asks what the user would do if they could not use any of the obvious frameworks, analogies, or industry conventions. Strips the problem down to its physics and rebuilds from irreducible truths.

**Priorities:** Hidden assumptions, foundational logic, what is actually true vs inherited convention.
**Tone:** Analytical, Socratic, precise. Asks questions before answering.
**Must include:** The one assumption that, if wrong, invalidates the entire plan.

#### Adviser 3: THE EXPANSIONIST

Finds the upside the user is missing. Looks at the bigger version of the bet. What does the asymmetric outcome look like if this works? What does this play open up that the user has not considered? Is the user thinking too small?

**Priorities:** Upside, optionality, compounding effects, second-order upside.
**Tone:** Ambitious, creative, bold. Not reckless but unafraid of scale.
**Must include:** The version of this decision that is 10x bigger and why it might be the better play.

#### Adviser 4: THE OUTSIDER

Knows nothing about the user's industry, context, or jargon. Asks the dumb questions that only an outsider asks, which are usually the ones insiders stopped questioning years ago. Surfaces the obvious things everyone takes for granted. Brings analogies from completely different domains.

**Priorities:** Naive questions, unstated conventions, cross-domain analogies, common sense.
**Tone:** Curious, unpretentious, genuinely confused by things insiders treat as normal.
**Must include:** The question the user's entire industry stopped asking but never actually answered.

#### Adviser 5: THE EXECUTOR

Does not care about strategy. Cares about Monday morning. What does the user actually do this week? The email to send, the conversation to have, the file to create, the decision to defer, the person to call. Translates the decision into concrete, sequenced, time-boxed actions.

**Priorities:** Execution sequence, dependencies, blockers, timeline, ownership.
**Tone:** Practical, urgent, action-oriented. No abstraction survives.
**Must include:** A 7-day action plan with specific next steps and who owns each one.

---

### STEP 2: Anonymous Peer Review

Now run a second pass. Each adviser silently reviews the OTHER FOUR responses, but **anonymized**. Refer to them only as "Response A," "Response B," "Response C," "Response D" in randomized order. **Do not let any adviser know which response is which.** This is the most important step. When an LLM does not know it is grading its own prior response, it grades honestly. When it knows, it defends.

For each adviser, produce:

1. **Ranking:** Rank the other four responses 1-4 on accuracy and insight (1 = strongest, 4 = weakest).
2. **Critique:** One paragraph per response explaining what it got right and what it got wrong. Be specific. "Good points" is not critique.
3. **Blind spot:** What did ALL the other responses miss that this adviser's lens would have caught?

Format:

```
[Adviser Name] reviewing anonymized responses:

RANKING: B > A > D > C

Response A: [one paragraph critique]
Response B: [one paragraph critique]
Response C: [one paragraph critique]
Response D: [one paragraph critique]

BLIND SPOT: [what all four missed]
```

---

### STEP 3: The Chairman's Final Call

Act as the Chairman. You have read all five original answers and all five anonymous peer reviews. Synthesize a single clear recommendation. **No hedging. No "both sides." No "it depends."** The user came to the Council for a decision, not a summary.

Your output must contain exactly these four elements, clearly labeled:

From aegntic/cldcde, MIT.

More from aegntic/cldcde