Brainstorm

Brainstorm — many distinct solutions, then pick

The failure mode this kills: the first plausible answer anchors the user, and N agents asked the same question converge into false consensus — three pitches that are one idea in three fonts. The job here is genuine divergence, then a clean pick. The description is the always-loaded when; this body is the how.

This skill is mode-based. Add modes here as the skill grows — never fork a sibling skill (kernel directive 6).

• Mode 1 · Solution Arena — breadth. Survey the known solution space: N

independent agents spread across orthogonal stances, then a clean side-by-side pick. The default for "what are my options / compare approaches / name this."

• Mode 2 · Invention Forge — depth. For when the known options all

disappoint and you want something that doesn't exist yet ("invent", "novel", "breakthrough", "first principles"). Heavyweight (multi-agent + web); not the default. Frames the problem to its invariants, then forces the model past its four default failure modes (below) toward a prior-art-grounded breakthrough.

Mode 1 · Solution Arena

0. Scope the problem

• Restate it in one sentence. If it's vague — no success criterion, no real

constraints — ask 1–2 clarifying questions FIRST. Diverse solutions to the wrong problem just waste the fan-out.

• Default N = 3 candidates. Honour an explicit count ("give me 5"); cap the

arena at 4 (see §4) — beyond that, generate more but shortlist before the pick.

1. Pick the diversity engine — ask the user

Call AskUserQuestion (single-select) offering these three engines. Recommend Fixed lenses as the default; if the user says "you pick," use it.

• Fixed lenses — each agent gets a distinct stance: Pragmatist (simplest

thing that ships) · Contrarian (invert the obvious approach) · Visionary (ideal world, ignore current limits). Add a 4th stance only if N>3.

• Per-problem dynamic angles — YOU read the problem and invent N

maximally-different angles tailored to it (e.g. for churn: pricing-lever / onboarding-lever / community-lever). No fixed personas.

• Distinct ideation methods — one technique per agent: first-principles

(rebuild from base truths) · analogy (how does another field / nature solve this?) · constraint-removal (what if money / time / compute were free?).

Whatever the engine, the assignments must be orthogonal: if two briefs would push toward the same mechanism, change one BEFORE spawning.

2. Generate — independent and parallel

• Spawn N subagents in one message (parallel Agent calls, general-purpose

type). Each gets: the scoped problem, ONLY its own lens/angle/method, and the charge: "You are one of N independent attempts. Do NOT hedge toward a safe middle — commit hard to YOUR angle. Your output is data for a picker, not prose for a human."

• Agents must not see each other — independence is the entire value. The only

deliberate exception is the divergence-guard respawn (§3).

• Require a structured pitch back from each: title (≤6 words) · core idea

(2–4 sentences) · why it's distinct · key risk / tradeoff · first concrete step. (If you'd rather scale past N≥5 or vet each candidate adversarially, run generate+guard as a Workflow returning these pitches as a schema — then do §4 in the main loop, since a workflow's agents have no channel to prompt the user.)

3. Divergence guard — enforce "truly unique"

• Read the N pitches. A collision is two pitches sharing the same **core

mechanism**, not merely similar wording.

• On collision: keep the stronger one; respawn the other with the colliding

pitches quoted and the brief "produce a solution that differs in MECHANISM, not just framing, from the following: ." Repeat at most once.

• Tell the user when you broke a collision ("agents 2 & 3 both landed on X —

regenerated 3"). Silent regeneration hides that divergence nearly failed.

4. Arena — side-by-side pick

• Present with AskUserQuestion, single-select, one option per candidate.

Put each candidate's full pitch in that option's preview so the UI renders them side-by-side — that side-by-side layout is the arena. Option label = the candidate title; description = a one-line hook.

• Plain OUTCOME language, not the mechanism's jargon (user taste — memory/

user-model.md, evidence 4). label/description/preview must name what each candidate does for the user and the one thing it found/changes — never the algorithm behind it. "What sets off what · almost nothing runs on its own", NOT "Reachability / transitive closure"; "Top-to-bottom layers", NOT "Tarjan SCC + longest-path layering". If you can't state an option in one sentence a non-engineer decodes, the pitch isn't finished — an undecodable arena earns "idk what these mean" and the pick stalls (happened 2026-06-19).

• Fallback: if previews don't render side-by-side (older client, or >4

candidates after shortlisting), present the pitches as a numbered list and ask which wins. The pick is what matters; the side-by-side layout is a nicety.

• Do not pre-rank or signal a favourite in the option text. The point is an

uncontaminated pick. If the user wants your read, give it AFTER they choose.

5. After the pick — offer follow-ups

Call AskUserQuestion (multiSelect) offering all three:

• Synthesize — graft the strongest ideas from the runners-up onto the winner

into one merged solution.

• Expand — turn the winner into a concrete plan / implementation steps.
• Re-brainstorm — run a fresh round with new lenses (loop to §1) if none

landed.

Return the result: the winner, or the merged / expanded artifact.

Mode 2 · Invention Forge

Use when the known solutions all disappoint and the user wants something that does not exist yet. The lever is not telling the model to "think harder" — weights are frozen (kernel). The lever is a process that routes around the model's four default failure modes:

1. Mean-regression — left alone it drifts to the safe middle → forced

orthogonal divergence (§1).

1. Self-unfalsification — it won't try to kill its own ideas → adversarial

gates (§3).

1. No prior-art grounding — it cannot certify "this doesn't exist" from

confidence → a real web search (§3).

1. Single-pass — it stops at first-order ideas → recombine + loop (§2, §4).

Read references/invention-forge.md for the per-vector agent briefs, the pitch schema, the three gate prompts, and the scorecard format before spawning.

0. Frame to invariants — the highest-leverage, most fallible step

• State the job-to-be-done as an OUTCOME, not the shape of today's solution.
• List the invariants: what ANY solution must satisfy (real constraints,

physics, the success criterion). This is what a candidate is judged against.

• List the inherited assumptions: conventions current solutions copy that are

NOT actually required. This list is the novelty surface — invention is dropping one. If you can't name any, the problem isn't framed yet.

• Pin the baseline: the boring, obvious, best-known solution. It is the

control every candidate must beat — a novel idea that loses to the baseline is theater, not a breakthrough.

• If invariants or the success criterion are vague, ask 1–2 questions FIRST (as in

Mode 1 §0). A breakthrough against the wrong criterion is wasted.

1. Diverge on invention vectors — parallel, independent

Spawn one general-purpose agent per vector in one message (reuse Mode 1's independence rule + its §3 divergence guard — Mode 2's own §3 is the funnel). Each gets the frame and ONLY its vector. Default the four below; swap one if a vector is dead for this problem.

• Assumption-drop — take the most load-bearing inherited assumption; design as

if it were false.

• Cross-domain transfer — find a structurally-isomorphic problem in a distant

field (nature, another industry, another scale/era) and transplant its mechanism. Invention is mostly recombination.

• First-principles rebuild — derive only from the invariants; ignore how it's

normally done.

• Constraint-to-extreme — push one constraint to 0 or ∞ (free compute, zero

latency, one user, infinite scale) and harvest what unlocks.

2. Recombine — cross-pollinate (deliberately breaks Mode 1's rule)

Read all vector pitches and graft their strongest fragments into 2–4 hybrid candidates. Mode 1 forbids candidates seeing each other; here the cross- pollination IS the point — that's why this is a separate mode. Carry the best pure vector pitches forward too; hybrids don't always win.

3. Adversarial filtration funnel — each candidate must clear ALL three gates

Run the gates per candidate (a Workflow pipeline scales this; see references).

• Feasibility kill-test — a FRESH skeptic agent tries to prove it can't work

or violates an invariant. Default-refute; the candidate survives only if the refutation fails.

• Prior-art gate — a fresh subagent runs WebSearch/WebFetch: does it already

exist? If yes it is not a breakthrough — drop it, or keep only the genuine delta vs prior art. Be honest in output: this proves "not found", never "doesn't exist".

• Dominance gate — does it actually beat the baseline (§0) on the success

criterion? If not, cut it. Tell the user the funnel arithmetic ("8 candidates → 3 survived: 2 killed on feasibility, 2 already exist, 1 lost to baseline").

4. Loop if thin

If <2 strong survivors, mutate the survivors and drop the NEXT inherited assumption, then re-run §1–§3 (loop-until-dry, max ~2 extra rounds). Log when you stop and why — silent capping reads as "explored everything" when it didn't.

5. Arena the survivors

Present survivors via Mode 1 §4 (AskUserQuestion, side-by-side, plain OUTCOME language — not mechanism jargon). Each carries an honest scorecard: novelty (vs the prior art actually found), feasibility risk, and margin over baseline. The user picks; then offer the Mode 1 §5 follow-ups (Synthesize / Expand / Re-brainstorm). Never pre-rank.

Rules

Both modes:

• The user picks — you don't pick for them. No "option 2 is clearly best" in

the arena. This is a generator, not a decision skill; the chosen solution is the user's, not a recommendation you then defend.

• If you can't make N orthogonal, propose fewer — distinct mechanisms beat

variations on one idea.

Mode 1:

• Independence is the value. Never let candidates converge by sharing context,

except the deliberate §3 respawn.

Mode 2:

• Generation stays independent; synthesis is deliberate. Vectors (§1) never see

each other — only the recombine step (§2) merges them, on purpose.

• Novelty is earned, never asserted. A candidate is "novel" only after the

prior-art gate (§3) fails to find it — and even then say "not found", not "new".

• The baseline is the control. Always carry the boring best-known solution

through to the scorecard; a candidate that doesn't beat it is not a breakthrough.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: GhostlyGawd
• Source: GhostlyGawd/recursive-harness
• License: MIT

Install and usage instructions live in the source repository linked above.