apw-brainstorm— agent skill

Use this skill when asked to brainstorm, ideate, generate options, or explore approaches before committing to one. It exists to fight one specific failure: a model defaults to its single highest-probability answer, so an unaided "give me ideas" produces a few near-identical, generic options, anchored to the first framing and presented uncritically. This skill forces genuine divergence, defers judgment, then converges honestly. It sits upstream of apw-research (verify the chosen direction) and apw-implement (build it). Follow the steps below strictly and in order.

Step 1 — Frame and re-frame the challenge

Before generating a single idea, write down what the ideation is actually for:

• Goal — what a successful outcome would achieve, in one sentence.
• Constraints — what is genuinely fixed, separated from what is merely assumed fixed. Mark each as hard or assumed.
• Success criteria — what a great idea would have to do to win, so convergence later has something to measure against.
• Re-framings — restate the problem 2–3 different ways (e.g. broaden it, narrow it, invert it). This deliberately breaks the anchor to the user's first phrasing before any ideating begins.

Present this framing — including the re-framings — and confirm it before going further.

Step 2 — Ask questions and set a divergence strategy (only if needed)

If anything about the framing is genuinely ambiguous, ask up to 5 targeted questions that would change the search space the ideation explores. Skip if the framing is already clear — do not ask for the sake of asking.

In the same round, decide the divergence strategy:

• Lenses — pick the idea-generation lenses you will apply in Step 3 so ideas don't all cluster (see the list there).
• Parallelism — independent agents generate decorrelated ideas, which is the strongest structural defense against mode collapse. If the challenge has separable angles (different user segments, domains, or sub-problems), tell the user the divergence could be split across parallel agents and ask whether to spawn them. If yes, give each agent the framing, a distinct lens or sub-angle, and instruct it to return raw ideas (no self-censoring). Synthesize them yourself in Step 5.

Step 3 — Diverge: generate widely, defer judgment

Generate ideas under an explicit discipline. No evaluation in this step — wild, half-formed, and "bad" ideas are all welcome; judging now is what collapses the output.

• Quantity target. Commit to a number up front (default: aim for at least 10 genuinely distinct ideas). Quantity is what forces you past the obvious ones.
• Apply distinct lenses so ideas span the space instead of clustering. Use several of:
• First principles — ignore how it's normally done; build up from what's actually true.
• Cross-domain analogy — how does another field solve the structurally similar problem?
• Inversion — "what would guarantee failure?", then flip each answer.
• Constraint removal — "if budget / time / a hard rule weren't real, then…"
• Constraint addition — "if you had one hour / one dollar / one file…"
• Persona shift — how would a different role, user, or competitor approach it?
• SCAMPER — substitute, combine, adapt, modify, put to other use, eliminate, reverse.
• Capture each idea with the lens that produced it and a one-line description.
• Collapse duplicates. If two ideas are the same idea reworded, merge them and push for a genuinely different one in its place. Reworded duplicates are mode collapse wearing a disguise.

Step 4 — Push past the first cluster

This is the core anti-mode-collapse gate. Stop and inspect the batch from Step 3:

• If the ideas are all variations on the safe, default answer, say so out loud and force several more that are contrarian, uncomfortable, or that break one of the assumed constraints from Step 1.
• The deliverable of this step is at least a few ideas that you would not have produced on the first pass. If every idea still sits in the obvious region, the divergence isn't done.

Step 5 — Converge: cluster and evaluate

Now, and only now, turn on judgment:

1. Cluster the ideas into themes so the shape of the space is visible, not a flat list.
2. Evaluate the strongest candidates against the success criteria from Step 1 in a small matrix with honest tradeoffs.
3. Kill weak ideas explicitly. Name the ones that don't survive and why. Presenting every idea as viable is sycophancy — the convergence is where you earn trust by cutting.

The full idea log, clustering, and evaluation matrix belong in the artifact (Step 7), not in chat. Do this work in service of the doc; don't paste it into the conversation.

Step 6 — Surface surprises and challenge the framing

If the ideation reveals that the real problem is different from the framing — the goal was wrong, a "hard" constraint dissolved, or the interesting question is elsewhere — stop and tell the user. Re-frame with them rather than brainstorming the wrong problem to completion.

Step 7 — Recommend, shortlist, and record

1. Converge to a ranked shortlist or a single recommendation with the reasoning behind it. Thirty ideas and no direction is not a finished brainstorm.
2. Say what each shortlisted idea would need next — typically a hand-off to apw-research to verify feasibility, or straight to apw-implement if the path is clear.
3. Write the result to your work item's `brainstorm.md` artifact — see apw-work-with-work-item-artifacts for where work-item artifacts live — using .claude/skills/apw-brainstorm/BRAINSTORM_TEMPLATE.md as the structure. Capturing the idea log (not just the winner) means the discarded options can be revisited instead of regenerated.

This artifact is the single detailed record — the full idea log, clusters, matrix, shortlist, and open questions live here. The chat (Step 8) gets a concise summary that points to it, never a second copy of the same content.

Step 8 — Concise summary

The detail is already in the artifact, so keep the chat result short — its job is to tell the user the headline and send them to the doc, not to reproduce it. Report only:

• The top recommendation (or the 2–3 leading shortlist items) in a line or two each.
• A one-line note that weaker ideas were considered and cut, and whether any assumed constraint fell.
• The path to the artifact, where the full idea log, matrix, and open questions live.

Do not paste the idea log, the clustering, or the evaluation matrix into chat — that is the duplication this step exists to avoid. Stay honest: if the real options are narrow, say so in the one line rather than padding the summary to look richer than the doc.