/brainstorm — Multi-Agent Brainstorming

Purpose: when someone is stuck on a product or business problem, surface the options they would have missed. The default result is 3 genuinely different paths they can act on this week, each with what it gains and what it gives up, which perspective it came from, and the devil's advocate's strongest pushback, plus one line on where the perspectives disagreed and one contrarian path that challenges their current direction. Every idea with its scores is available on request ("full table").

How the options are found: 5 subagents run in parallel, each with a different fixed lens, and their ideas are then converged with DVFN scoring. This borrows the "individual divergence → group convergence" structure of Google Design Sprint and implements it with isolated agents.

Why this structure

Research on human group brainstorming identifies three causes of productivity loss — production blocking (forgetting ideas while waiting your turn), evaluation apprehension (self-censoring), and social loafing (coasting on others). Running isolated agents in parallel and converging only afterward is a design inspired by that research. It is not proof that the output is better: five agents running the same model can still converge on similar ideas, which is why each one gets a distinct fixed lens, and why the paths are chosen to differ in mechanism.

Input

/brainstorm <problem> — already tried: <…> — limits: <…> — goal: <…>

Example: /brainstorm week-one signups for our launch are flat. Already tried: a Product Hunt post and a waitlist email. Limits: no paid ads, 1 marketer. Goal: 500 signups in week one.

Language: write the whole report in the language the user wrote the topic in.

Output mode:

• Paths (default) — the short report in [5].
• Full table — if the user asks for "full table" (in the topic or after a Paths report), produce the full-table report in [5]. If they ask after a Paths report, reuse the ideas and scores from [3]; do not run divergence again.

Procedure

[0] Context check (one turn)

Before diverging, ask the user one short message covering whatever is missing from the topic:

• Who is the target (customer / user / audience)?
• Goal — what outcome counts as success?
• Constraints — budget, time, team, things that are off-limits
• What have you already tried or considered? (optional — say it is optional)

Ask only for what is missing. If the user says "skip", says that's all they know, or the topic already answers these, proceed and mark every assumed part as "(assumed)" in the report. Never invent context silently.

Constraints are hard. Budget, team, time, and off-limits items stated here are limits, not preferences. An idea that breaks one is dropped or adapted to fit in [3], and the report says so in one line.

Ask in plain text and wait for the reply. If no reply is possible (for example, a non-interactive run), proceed with marked assumptions.

[1] Reframe as How Might We

Rewrite the topic as 1–3 open "How might we …?" questions. Confirm with the user in one line (if they correct it, use the corrected version). If the user has asked you to proceed without stopping, or no reply is possible, state the HMW you are using and continue.

[2] Parallel divergence — 5 subagents

Call the Agent tool (named Task in older Claude Code versions; use the general-purpose agent) 5 times in a single message so they run in parallel. Give each agent the same HMW question, the context from [0] (including the hard constraints and the user's already-tried list), and exactly one fixed lens:

Each lens owns one common reason ideas fail, and the lenses diverge independently so they don't anchor on each other.

Tell each agent the question its lens owns.

Shared instructions for every agent:

• 5–7 ideas. Defer judgment (IDEO rule) — quantity first. Wild ideas are fine inside the hard constraints
• Do not repeat anything on the user's already-tried list. A variant of a tried idea is allowed only if it starts with one line: New vs. tried: <what is different from what they already did>
• Ignore the other agents entirely (independent divergence)
• Each idea: a one-line title + 2–3 sentence description
• Tool limit: divergence only. No file edits, no external messages

Lens 4 also ends with 3–5 one-sentence failure scenarios for the most likely approaches to this problem (for example "a new event fails because…"), so convergence can attach real pushback to each path.

Lens 5 output is a Synthetic persona response. Keep that label in the report and in anything passed to other tools; never present it as real user research.

[3] Converge — filter, dedup, DVFN scoring

• Remove ideas that duplicate the user's already-tried list. Keep a variant only if its "New vs. tried" line names a real difference
• Apply the hard constraints. Drop each idea that breaks one, or adapt it to fit and note the adaptation. Count both for the report. Ideas cut only for a low score are not constraint drops and do not go in that count
• Merge the remaining raw ideas by similarity. Record which lens(es) each merged idea came from
• For each idea, note the lens 4 failure scenario or counter-argument that bears on it most directly, if any
• Note where two lenses pulled in opposite directions on the same point (one proposes something, another argues it is the problem). Keep what each actually said. If no two lenses clearly disagreed, record that
• When merging, do not compress an agent's description into one line. Keep the substance (why it matters, concrete examples) as material for [5]
• Tag each idea with its mechanism — the route by which it would move the goal (for example: bring in a new audience, shift existing customers to a different time, give a reason to come on a specific day, borrow another business's traffic, change the offer itself)
• Score each idea 1–5 on four axes — always write the full names:
  • Desirability — does the target actually want it?
  • Viability — can it make money / scale?
  • Feasibility — can it be built with the stated constraints?
  • Novelty — is it differentiated?
• Give each idea a confidence level — exactly one of High, Medium, Low — for how much the scores rest on evidence vs. guesswork
• Synthetic persona responses (lens 5) are hypotheses, not evidence. Score rationales never use lens 5 or agreement between lenses as a reason (lenses share one model, so agreement is not evidence). Score on the idea itself and the stated context
• Whenever lens 5 content appears anywhere in the report, use exactly this form: Simulated user (synthetic): "<quote>", then on the next line Hypothesis: <what it suggests, to be tested>
• Sort by total score

[4] Choose the paths (Paths mode)

• 3 paths: go down the sorted list and take the highest-scoring idea for each mechanism until you have 3 paths with 3 different mechanisms. Ideas that share a mechanism may be folded into one path. Three variants of one idea (three events, three discounts) are not three paths
• 1 contrarian path: an option that reverses something the user is currently doing or assuming — the direction their tried list points in, the goal metric, or the audience they are aiming at. It usually comes from lens 4 or from a cut idea. It is not a fourth ordinary path
• Every "this week" first step must fit the hard constraints
• The contrarian path's check tests the contrarian move itself, so its pass line can only be met if the user actually does the reversed thing

[5] Report

Plain language: if the audience is not a domain expert, gloss every technical term or acronym in place, in the form "plain explanation (term, ACRONYM)" — e.g. "a hidden factor that muddies the real cause (confounder)".

Paths report (default) — about 30–45 lines

Keep it short: no paragraph under a path name — only the seven fields, each one sentence. Address the user as "you". Keep lens 5 quotes out of the Paths report; state key assumptions as assumptions. (Lens 5 detail belongs in the full table, in the Simulated user (synthetic) form.)

The visible perspective layer is what plain chat cannot give, so it is required:

• From: the lens(es) whose raw ideas the path came from, each followed by its job in parentheses so the reader sees what that lens checks: PM (fast test) · Strategist (pays and lasts) · Designer (ease of use) · Devil's advocate (where it breaks) · Simulated customer (wanted?, synthetic). Join several with "+", e.g. "Designer (ease of use) + PM (fast test)". Lens 5 always carries "synthetic"
• Strongest pushback (Devil's advocate): one sentence, the most serious reason this path could fail, taken from what lens 4 actually wrote. If lens 4 said nothing that applies to this path, write it yourself and label the field Strongest pushback (AI assumption): instead
• Where the perspectives disagreed: one line naming two lenses and the concrete point they pulled apart on, from their actual outputs. If one side is lens 5, write it as "Simulated customer (synthetic)" and state its side as a hypothesis, without a quote. If no two lenses clearly disagreed, say so; never invent a disagreement

## HMW: <reframed question>

**Context:** target · goal · constraints · already tried (mark anything assumed as "(assumed)")
**Fit to your limits:** <one line, only ideas that broke a hard constraint — e.g. "2 ideas dropped (needed renovation), 1 adapted to existing tables"; or "nothing dropped">

### Path 1 — <one-line name>
- **From:** <lens name(s)>
- **This week:** <a concrete first step, doable this week within the constraints>
- **You gain:** <what this gets you>
- **You give up / risk:** <the real cost — money, staff time, customers, brand, or what you stop doing>
- **Strongest pushback (Devil's advocate):** <the most serious reason this path could fail> (or `Strongest pushback (AI assumption):` if lens 4 didn't address it)
- **Key assumption:** <the belief that, if wrong, kills this path>
- **Next validation experiment:** <the cheapest test of that assumption> — **Pass = <a specific, observable result>**

### Path 2 — … (same fields)
### Path 3 — … (same fields)

**Where the perspectives disagreed:** <Lens A> <its position> vs. <Lens B> <its position>, on <the concrete point> (or: "The lenses did not clearly disagree this time.")

### Contrarian path — <one-line name>
- **From:** <lens name(s)>
- **It flips:** <the current direction or assumption it reverses>
- **Why it might be right:** <1–2 sentences>
- **Cheapest check:** <what to do> — **Pass = <observable result>**

### The 3 paths side by side (AI assumption — qualitative judgment, not measured)
| | Path 1 | Path 2 | Path 3 |
|---|---|---|---|
| Cost / effort | | | |
| Speed to signal | | | |
| Upside | | | |
| Main risk | | | |

Want every idea with scores? Say "full table".

Full-table report (on request)

Always put the full-name legend above the table — no bare initials. Every top idea gets a one-line summary plus a 2–4 sentence detail; a single compressed line is not enough.

## HMW: <reframed question>

**Context:** target · goal · constraints · already tried (mark anything assumed as "(assumed)")
**Fit to your limits:** <one line, as in the Paths report>

> Score legend: **Desirability** · **Viability** · **Feasibility** · **Novelty** — 1–5 each. These are the assistant's judgments, not measurements.

### Top ideas at a glance
List the top 5 (no more). Every idea in this table gets a full entry in "Top ideas in detail".
| Rank | Idea | Desirability | Viability | Feasibility | Novelty | Total | Confidence | Source lens |
|---|---|---|---|---|---|---|---|---|

### Top ideas in detail
For each top idea:
- **[Idea]** — *Summary*: one line
  - *Detail*: 2–4 sentences, keeping the agents' concrete content and examples
  - *Score rationale*: Desirability N / Viability N / Feasibility N / Novelty N — one line each
  - *Key assumption*: the belief that, if wrong, kills this idea
  - *Next validation experiment*: the cheapest test of that assumption — what to do, and **Pass = <a specific, observable result>** (an experiment without a pass line is incomplete)

### Wildcards — cut but interesting (1–2, summary + detail)
### Assistant's recommendation (state that this is a subjective judgment)

Final check — run this on the finished report before sending it

Re-read the finished report text itself (do not answer from memory) and fix it until every answer is "yes".

Both modes

1. Compare every path or idea with each item on the user's already-tried list. Is none of them the same thing under a new name? Where one is a variant, does it say in one line what is new? Remove anything that only repeats the list.
2. Does every first step, path, and idea fit the hard constraints (budget, team, time, off-limits)? Does the "Fit to your limits" line say what was dropped or adapted?
3. Read every sentence that draws on lens 5 or on lens agreement. Does any of them state, as fact, what users want, feel, need, or confirmed, or say the simulated user "was right"? Rewrite each one into the Simulated user (synthetic) + Hypothesis: form. Does any score rationale mention lens 5 or lens agreement? Remove it. (Checking for specific words is not enough. Check what each sentence means.)
4. Take every specific detail that came only from lens 5 (prices, quotes, situations, feelings) and look for it everywhere in the report. Wherever it appears without the Simulated user (synthetic) form, rewrite it or remove it. Do the same for claims about what the target group thinks or does that have no source: label them as hypotheses.
5. Does every validation experiment and contrarian check contain "Pass =" with an observable result? Can the stated measurement method record every condition in the pass line, and is there a baseline to compare against? If not, change the method or the pass line.

Paths report

6. Do the 3 paths work through 3 different mechanisms? Test: if Path 1 worked, would that tell you almost nothing about whether Path 2 or 3 would work? If two paths are the same move in different clothes, replace one.
7. Does each path have both a gain and a give-up, and is the give-up a real cost (money, staff time, customers, brand, or something the user stops doing) rather than "none" or "takes effort"?
8. Is the contrarian path present, and does it actually reverse something the user is currently doing or assuming, rather than being a fourth ordinary path? Does its check test that reversed move (not "do nothing"), and does it speak to the user as "you"?
9. Is the comparison table labeled as an AI assumption, qualitative, and does the report end with the "full table" line?
10. Does every path and the contrarian path have a From: line, with each lens followed by its job in parentheses? Open the raw lens outputs: did that path's ideas actually come from the lenses named? Is lens 5 written with "synthetic"? Naming lens 5 in From: does not relax checks 3 and 4: no lens 5 quote or lens-5-only detail appears in the Paths report as fact.
11. Does every path have a Strongest pushback line of one sentence? If it is labeled "Devil's advocate", can you point to the lens 4 sentence it comes from? If you can't, relabel it Strongest pushback (AI assumption):.
12. Is the Where the perspectives disagreed: line present between Path 3 and the contrarian path? Can you point to the sentence in each named lens's raw output that takes each side? If you can't, rewrite it with a disagreement you can trace, or say the lenses did not clearly disagree. Never invent one.
13. Count the lines of the Paths report. Is it 45 lines or fewer, with no paragraph under any path name and one sentence per field? Does the "Fit to your limits" line list only constraint breaks, not ordinary low-score cuts? If not, cut.

Full-table report

14. Is every Confidence value exactly High, Medium, or Low? (Values such as "Medium-High" or "Med" are not allowed. Pick one.)
15. Does the top-ideas table have 5 rows or fewer, each with a detail entry?

Guardrails

• ❌ No filtering or self-censoring during divergence beyond the hard constraints and the already-tried list — convergence happens only in [3]
• ❌ No DVFN score without a one-line rationale
• ❌ No bare initials (D·V·F·N) — always full names
• ❌ No one-line-only ideas in the full table — summary + 2–4 sentence detail
• ❌ Do not present simulated-user output as real user evidence
• ✅ The 5 agents must be dispatched in one message — parallel, independent divergence is the point
• ✅ Label scores, the comparison table, and the recommendation as the assistant's judgment, separate from sourced facts