# Skill: Assemble a four-role AI council to evaluate a hard decision

> Lets an AI coding/design agent break anchoring bias by consulting four fresh, adversarial perspectives on a genuine fork-in-the-road decision -- applies whenever there are multiple live options and the cost of a wrong choice is real.

Source: sodigi·learn -- vibecoding/sovet-iz-chetyreh-ii · https://sodigi.io/learn/vibecoding/sovet-iz-chetyreh-ii

## When to use
- A decision has several plausible options with no clear winner (architecture choice, "ship now vs. polish," build vs. buy).
- You suspect the current conversation is biasing the AI toward agreeing with what you already want (anchoring).
- A mistake is costly or hard to reverse and you want an honest map of pros and cons first.
- You need to see the strongest objection surfaced, not buried.

## Core rules
- Use the council only for genuine forks in the road -- not for factual questions, execution tasks, bug hunts, or step breakdowns. Those have better-suited tools.
- The council has exactly four roles: **Architect** (correctness, reliability, long-term soundness), **Skeptic** (do we even need this?), **Pragmatist** (fastest path to real value), **Critic** (what will break, what is at risk).
- Three of the four voices must receive only the question and minimal context -- NOT the full chat history. This is the vaccine against nodding along.
- The summary must surface the **strongest disagreement** prominently, not smooth it over.
- The council delivers a map of the disagreement; the human presses the final button.
- The council costs four rounds of reasoning -- reserve it for decisions that justify the expense.

## Procedure
1. Identify that the user is at a genuine decision fork with multiple live options.
2. Extract: (a) the question in one sentence, (b) minimal context needed to answer it cold.
3. Spawn the four roles. The current conversation handles the Architect. The Skeptic, Pragmatist, and Critic each receive only the question + minimal context, no prior chat.
4. Generate each opinion separately, labeled by role, from the relevant angle.
5. Pull together a closing summary: where all four agree, where the sharpest disagreement is, and a personal recommendation -- while explicitly leaving the final decision to the human.
6. Do not manufacture false unanimity. If voices conflict, keep the conflict visible.

## Ready-to-use prompt
```
Assemble a council of four voices on my decision.

Question: [state the decision in one sentence]
Context: [2-4 sentences of relevant facts; omit the rest of the conversation]

Give four separate opinions, each in its own labeled role:
- Architect: how to do it correctly and durably
- Skeptic: whether we even need this, and whether there is a simpler way
- Pragmatist: how to get the most value as fast as possible
- Critic: what could break and what are the real risks in each option

End with a short synthesis:
- Where all four agree
- Where the sharpest disagreement lies
- Your own recommendation
- Note that the final decision is mine.
```

## Pitfalls
- Calling a council for trivial or factual questions -- waste of time and tokens; just ask directly.
- Giving all four voices the full chat history -- they will pick up your mood and nod in chorus, defeating the whole purpose.
- Expecting a single "right answer" -- the value is in the visible disagreement, not in unanimity.
- Burying the strongest objection in the synthesis instead of highlighting it.
- Delegating the final choice to the AI -- the council informs the decision, it does not make it.
- Confusing the council with other tasks: finding bugs, writing code, and planning steps are different tools.