Decision council

Five advisors, one answer. The value is the disagreement: a panel that nods along has told you nothing.

Run it on the question the user asked. No question given? Ask for the dilemma first; don't council whatever was last in the conversation.

When it is worth it

Worth it: hard to undo, costs real money or time, wrong-and-quiet is expensive: architecture, pricing, build/drop, risky release, "should we rewrite this?".

Not worth it: anything with a checkable answer. "Does this work?" → run tests. "Which is faster?" → measure. The council is for judgement, not facts.

The panel

Each advisor gets 3–4 sentences. For smaller questions use 3 advisors (Contrarian, First Principles, Operator).

Advisor	Job	Must not
Contrarian	Name what will fail, the load-bearing unexamined assumption, what going wrong looks like.	Be negative about everything; attack a weak version.
First Principles	What is really being decided? Is the framing wrong? What would you do with no sunk cost?	Re-derive the status quo and call it insight.
Expansionist	The bigger play behind the small one; the uncosted upside.	Wish. Every upside must be reachable from here.
Outsider	A worldview from another trade or a customer who has never heard of this.	Be a second Contrarian or Expansionist.
Operator	What to actually do: first step, order, owner, cost, what to cut.	Give a plan that only works if everything goes right.

Process

1. Views. Each advisor speaks once, in order, under a heading.
2. Cross-examination. Name at least one real conflict and who is more nearly right, and why. If someone was convinced, say what changed their mind. If all agree, say so and say what would make them all wrong together.
3. Chair. One recommendation:
   • the call in 1–2 sentences, first;
   • what it costs and what is given up;
   • the strongest objection left standing, and why proceed anyway;
   • next steps in order, small enough to start;
   • the signal that means "we were wrong".

Rules

• The chair may overrule the majority. Four agreeing is not a vote.
• Disagree with the user when the panel does. "Don't do this" goes first.
• No hedging in the call: no "consider", "it depends", or three options handed back.
• Facts beat opinions. If a point can be checked (in the code, docs, data), check it first, or send a scout to find out, then give the panel the finding.

Output

See references/output-template.md. To keep the decision, write it as an ADR (adr) with the chair's call as the decision.

Customize for your project

Swap in advisors that fit your domain (e.g. a Customer, a Security Lead), keep the count at 3–5, and keep the "must not" column.