consensus-plan

name	consensus-plan
description	Planner → Architect → Critic deliberation loop — produces a formally validated ADR
version	0.1.0
level	3
triggers	["consensus plan","deliberate on this","get a second opinion on the plan","validate this plan","/consensus-plan"]
pipeline	["consensus-plan","plan-first"]
context_files	["context/user.md","context/project.md","context/decisions.md"]
steps	[{"name":"Planner Draft","description":"Invoke the planner agent to produce an initial phased implementation plan"},{"name":"Architect Steelman","description":"Invoke the architect agent to find the strongest objection to the plan"},{"name":"Convergence Check","description":"Does the objection have a clear resolution? If yes, revise and proceed. If fundamental, loop back."},{"name":"Pre-Mortem (high-risk only)","description":"For HIGH risk plans — identify 3 specific failure scenarios before finalizing"},{"name":"ADR Output","description":"Write the formal Architecture Decision Record to context/decisions.md"},{"name":"Confirmation Gate","description":"Present the ADR. Wait for explicit approval before any implementation begins."}]

Consensus-Plan Skill

A structured deliberation protocol that forces a plan through challenge before it is approved. The output is a formally validated ADR — not just a plan the author thinks is good, but a plan that survived its strongest objection.

What Claude Gets Wrong Without This Skill

Without deliberation, a plan reflects only the assumptions of the entity that created it. Those assumptions are invisible to the planner — they only become visible when someone with a different perspective examines the same problem. The Planner → Architect → Critic loop exists to surface blind spots before implementation, not during it.

The Deliberation Loop

Round 1 — Planner Draft

Invoke the planner agent with the task description. The planner produces a structured plan (phases, steps, risks, parallel opportunities) in the standard format.

Do not modify the plan before passing it to the architect.

Round 2 — Architect Steelman

Invoke the architect agent with:

• The planner's draft
• This instruction: "Find the strongest objection to this plan. Steelman it — argue for the objection as if you believe it. Then state the minimum modification to the plan that would resolve it."

The architect must produce:

1. The strongest objection (not a list — the single strongest one)
2. The steelman argument for that objection
3. A specific modification that resolves it

If the architect cannot find a meaningful objection, it must say so explicitly: "No material objection found. Plan is sound." This is a valid outcome — not a failure to engage.

Round 2.5 — Anonymous Peer Review

After the architect steelman, invoke a third reviewer (researcher agent) with:

• The planner's draft
• The architect's objection and proposed modification
• This instruction: "Review this plan as if you have no stake in either side. You do not know who wrote the plan or the objection. Find the one assumption that both the planner and architect have accepted without questioning."

The peer reviewer must produce:

1. The shared assumption (one, not a list)
2. What would have to be true for that assumption to be wrong
3. A recommendation: proceed / revisit assumption before continuing

This step prevents anchoring: both the planner and architect may share a blind spot that only an unprimed reviewer can surface.

If the peer reviewer finds no unquestioned assumption, it must say so explicitly.

Nudge mechanism: After peer review, issue a forced challenge to the planner: