AGENT SKILL

aac-forge-idea

Domain-agnostic Socratic idea pressure-testing. Use when the user has a half-formed idea and wants to harden or kill it cheaply, one question at a time. Supports an adversarial attack mode and optional persona rooms resolved from the installed agent roster. Menu code FI.

License 许可
MIT
最近更新
2026-08-29
Stars
★ 3
Forks
0

01 前言 Front Matter

name
aac-forge-idea
description
Domain-agnostic Socratic idea pressure-testing. Use when the user has a half-formed idea and wants to harden or kill it cheaply, one question at a time. Supports an adversarial attack mode and optional persona rooms resolved from the installed agent roster. Menu code FI.

03 SKILL.md 正文

Forge Idea

Pressure-test a half-formed idea through Socratic, one-question-at-a-time interrogation until the idea hardens, proves out, or dies cheaply. The point of forge-idea is not to generate more ideas — that is brainstorming's job. The point is to take one idea and see if it survives contact with a hostile interlocutor.

When to Use

  • The user says "I have an idea" / "what do you think of X" / "should I build this" / "forge this" / "FI" / "pressure-test this"
  • The user has a half-formed concept that needs sharpening before any planning
  • A brainstorming session produced a candidate the user wants to validate
  • The user is at the very start of a project and wants to discover the real shape of the problem

Do NOT use for:

  • Bulk ideation (use aac-brainstorming)
  • Implementation details (use aac-create-architecture or aac-spec)
  • Document review (use aac-advanced-elicitation)

Modes

  • Socratic (default) — one question at a time, deepening. The agent asks; the user answers; the agent reflects and asks the next question that targets the current weakest link.
  • Adversarial — switch on with --attack or by answering "Attack: on" to the mode prompt. The agent's questions become hostile: pre-mortem, red team, kill shot. The user is forced to defend. Best for "I think this is great, prove me wrong."
  • Persona room — switch on with --room or by answering "Room: on". Each question comes from a different persona resolved from the installed agent roster (PM, Dev, QA, Architect, etc.). Multiple agents, in turn, attack the idea from their angle. Best for cross-functional validation.
  • Combined — --attack --room is the strongest setup: hostile questions from rotating personas. Use sparingly; it is exhausting.

The Loop

  1. The user states the idea in one or two sentences. The agent does not paraphrase back — paraphrasing is friction.
  2. The agent asks ONE question. Waits for the answer. Never bundles two questions into one round.
  3. Based on the answer, the agent:
    • Notes a residue (a fact, assumption, or contradiction the answer revealed) — kept in conversation memory, not persisted unless the user opts in
    • Picks the next question — the one that most threatens the idea's weakest link right now
    • Asks it. Waits.
  4. After N rounds (default 7, configurable with --rounds N), the agent offers a verdict:
    • HARDENED — the idea has survived the pressure. Move to brief or implementation.
    • DEAD — the idea collapsed. Note why. Do not revive unless the user asks.
    • UNRESOLVED — the user is still defending gaps. Offer to continue, switch modes, or stop.

The verdict is honest, not optimistic. A "HARDENED" verdict is rare and should be earned.

On Activation

  1. Parse arguments — check for --attack, --room, --rounds N, --headless.
  2. Load config from {project-root}/_bmad/core/config.yaml:
    • user_name for greeting
    • communication_language for all speech
  3. Resolve the agent roster (only if --room is set) by running:
    python3 {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root} --key agents
    
    Pick 3-5 personas with the most orthogonal expertise to the idea. Each will ask one question per round in turn.
  4. Welcome the user and surface the modes:
    • "Socratic (default) — I ask one question at a time. You answer. We go deeper."
    • "Adversarial (--attack) — I try to kill the idea. You defend."
    • "Persona room (--room) — Different agents take turns questioning. Useful for cross-functional check."
    • "Combined (--attack --room) — Hostile questions from rotating personas. Intense."
  5. Ask for the idea: "State it in one or two sentences. Don't polish."

Question Crafting

The agent's discipline:

  • One question per round. No "and". No comma-spliced pairs. If two questions feel necessary, pick the one that closes more risk.
  • Target the weakest link, not the strongest. The strongest part of the idea is already defended; pressure-testing it wastes a round.
  • Concretize. Replace "users" with a specific person. Replace "fast" with a number. Replace "easy" with a scenario. Vague answers reveal vague ideas.
  • Force a stance. "What would have to be true for this to work?" beats "do you think this will work?"
  • Track what you have learned. The residue list grows. When the user contradicts an earlier answer, name it.

Common question types to rotate through:

  • Concretize — "Give me a specific user. Not a persona, a person."
  • Why-now — "What changed in the world that makes this possible or necessary now?"
  • Pre-mortem — "It's 18 months from now and this failed. Why?"
  • Substitution — "What if you removed the central mechanism — does anything survive?"
  • Adjacency — "What is the closest existing thing? Why doesn't it just do this?"
  • Cost — "What does this cost if it succeeds? Not money — attention, focus, optionality."
  • Counter-user — "Who loses if this wins?"
  • Unstated assumption — "You just claimed X. What if X is wrong?"

Output (Optional)

At the verdict step, the agent offers to write:

  • A brief — 1-page summary of the hardened idea: problem, who, why now, shape, next concrete step. This brief feeds aac-spec (full spec) or aac-quick-dev (small change). Format matches the upstream aac-product-brief schema.
  • A residue log — append-only log of the rounds, the questions, the answers, the assumptions revealed. Useful as input to a future forge-idea session, or to seed a future aac-advanced-elicitation review.

The brief is opt-in. The user may want to keep the idea private until they have shared it elsewhere first.

Halt Conditions

Halt immediately if:

  • The user says "stop" / "kill" / "done" / "that's enough" — any natural phrasing
  • The user has nothing substantive to add (twice in a row, the second time is the halt)
  • The idea has clearly died (the user agrees with the kill assessment)

When halting, do not pad with summaries unless the user asks. A clean exit is the point of the skill.

Headless Mode

--headless runs the loop non-interactively:

  1. The idea is read from {project-root}/forge-input.md (or the first user message).
  2. The agent runs all N rounds in one pass, asking questions and answering them itself as a hostile user-of-one (best-effort — headless is weaker than interactive for forge-idea because the user is the source of new information).
  3. The verdict and brief are written to {project-root}/forge-output.md.
  4. The residue is written to {project-root}/forge-residue.md.

Headless is useful for overnight batch processing of idea backlogs, not for the idea you actually care about.

Success Metrics

  • The user states the idea clearly within 2 sentences (no prompt-induced bloat)
  • One question per round, never a barrage
  • Each question targets the current weakest link, not the strongest
  • The verdict (HARDENED / DEAD / UNRESOLVED) is honest, not optimistic
  • The optional brief is concise (1 page) and actionable (has a next concrete step)
  • The residue log is append-only and never edited

Failure Modes

  • Asking two questions in one round ("What about X? And also Y?") — wastes a round
  • Paraphrasing the user's idea back to them — friction
  • Generating ideas instead of testing one — that's brainstorming
  • Marking HARDENED too early — undermines the skill
  • Marking DEAD before the user has defended — paternalism
  • Continuing past UNRESOLVED without checking in — the user may be stuck

04 同仓库更多 Skills