speckit-roundtable

User Input

$ARGUMENTS

Why this exists

A single reasoner resolves trade-offs too early and too quietly. It picks the option it thought of first, then writes the justification. BMAD's answer is to run the discussion with several personas in character; the value is not the role-play, it is that the disagreement gets written down instead of being averaged into one confident paragraph.

This is deliberately not a subagent fan-out. The existing reviewers (spec-challenger, code-reviewer, spec-reviewer) already give independent readings of an artifact that exists. This is for a decision that has not been made yet, where the cost is in the trade-off and not in the evidence.

Seat the table

Four positions by default. Use the ones with something at stake in this decision, and add a domain-specific one when it is obvious.

Position	What it argues for	What it will sacrifice
The maintainer	the version still legible in a year; fewer moving parts	short-term convenience, cleverness
The user	the behaviour they can predict from the outside	internal elegance
The sceptic	the smallest thing that could work; does this need to exist?	completeness, future-proofing
The attacker	what happens under malformed input, concurrency, interruption, scale	ergonomics

Run it

1. State the decision as a choice, with at least two named options. "Should we improve X" is not a decision; "A or B, and what we lose either way" is.
2. One round, each position in turn. Two to four sentences. Each must name the option it backs and the cost it accepts. No position is allowed to agree with everything.
3. One rebuttal round, only where positions actually conflict. Skip it where they do not — a manufactured argument buries the real one.
4. Name the disagreement that survives. If every position converged, say so plainly; that is a real and useful result, not a failed exercise.

Output

Decision: <A or B>

Maintainer: <position, cost accepted>
User:       <…>
Sceptic:    <…>
Attacker:   <…>

Unresolved: <the disagreement that survived, or "none">
Recommendation: <option>, because <the consideration that outweighed the rest>
Rejected: <option>, which would have been right if <condition>

That last line is the point. A recommendation with no stated condition under which it flips is a preference wearing a suit.

Boundaries

• It decides nothing and writes no file. Hand the result to /speckit-specify, /speckit-correct-course, or the user.
• It does not spawn subagents. Constitution II: an orchestration nobody asked for is cost, not rigour.
• Four positions, one round, one rebuttal. A roundtable that runs long stops being read, and an unread transcript is worse than no transcript.