dialectic-brainstorming

name

dialectic-brainstorming

description

Brainstorm a consequential, no-ground-truth question with a panel of diverse minds, then resolve it dialectically. A casting step derives orthogonal thinking-axes for the question (each a quality to maximize) and casts one persona per axis; the personas propose and cross-critique concrete candidates while a neutral Synthesizer prunes, recombines across axes, and holds a frontier — returning a short list, not one winner. Use for naming (products, tools, projects), positioning, or any choice of consequence with no objectively correct answer where you want many genuinely different perspectives, not one. Triggers: /dialectic-brainstorming, "brainstorm this with a diverse panel", "what should we name…", "give me a shortlist from multiple perspectives".

dialectic-brainstorming — diverse minds diverge, dialectic converges

A sibling of dialectic. Where dialectic runs three structural roles (thesis/antithesis/ synthesis) to drive one question to a single answer, this skill runs N diverse persona-minds that diverge through different backgrounds, then a neutral Synthesizer converges them to a short list via dialectical sublation. Diversity is the engine of divergence; dialectic is the engine of convergence.

Use it for consequential questions with no ground-truth answer — naming, positioning, "which of these directions", any judgment call where you want many genuinely different minds rather than one voice averaging over itself.

Mechanism in one line

Cast orthogonal quality-axes → each persona maximizes its quality with concrete candidates → they cross-critique → the Synthesizer prunes weak candidates using the critiques and recombines across axes (crossover) onto a frontier → return the surviving short list.

Two modes — pick before casting

Each subagent spawn costs ~27k tokens flat (the agent's system prompt + tool schemas dominate; the brainstorming text is noise next to it). So spawn count is the cost, and the mode is a cost/independence trade:

• Batched (DEFAULT — low-to-mid-stakes). A single subagent role-plays the whole panel in one call: give it all N axes/personas and have it return each persona's PROPOSE+CRITIQUE independently and divergently (instruct it explicitly: argue each persona on its own axis, do not let them agree or blur). Then one Synthesizer spawn. ≈ 2 spawns/round (~55k) regardless of N. Saving vs isolated = the (N−1) persona-spawns you skip each round — measured ~2× at N=3, scaling to ~3× at N=5 (not more; the Synthesizer spawn is unavoidable). You lose true isolation (one model holds all personas at once) but keep the structure, and in testing the personas stayed distinct and the critiques stayed axis-partisan.
• Isolated (HIGH-STAKES only). The full protocol below — one fresh subagent per persona per round, genuine independence (no persona ever sees another's reasoning). Costs ≈ N+1 spawns/round; an N=3, 2-round run ≈ 16 spawns ≈ 430k tokens. Use only when the decision is consequential enough to pay for real independence.

Default to batched. Escalate to isolated only when the user says it's high-stakes or explicitly wants maximally independent minds.

1. Cast once (the divergence skeleton)

Before any brainstorming, analyze the question and derive orthogonal thinking-axes relevant to this question — 3 by default, 4–5 only if the question genuinely spans that many distinct qualities. Each axis is defined by a quality it optimizes. For "name a CLI tool": memorability, semantic fit/meaning, distinctiveness/ownability. For a strategy call, entirely different axes.

Then cast one persona per axis — a single-quality maximizer with a concrete, question-fit background and personality (e.g. a poet for aesthetics, a systems engineer for the technical axis, a cognitive scientist for human factors, a sharp contrarian for the skeptic axis). The background is invented per question, not fixed; the axis is the skeleton, the persona is the flesh.

• Axes must be orthogonal — if two collapse into the same lens, merge them and add a real third.
• Orthogonal axes do NOT guarantee orthogonal outputs. Independent personas can still converge on the same obvious answer (e.g. every persona proposing the same metaphor). This is expected for questions with a strong attractor. Tell each persona to differentiate from the obvious, have the Synthesizer de-duplicate convergent proposals, and treat a full collapse as signal that the attractor may genuinely be the answer — not as failure.
• The roster is frozen after casting. Cast once, run to convergence. (Adaptive re-casting — swapping a dead axis for a live one mid-run — is a deliberate v2 extension, not v1.)
• Emit the roster up front so the user can watch and redirect: AXES: memorability⟨the rapper⟩ · meaning⟨the etymologist⟩ · ownability⟨the trademark lawyer⟩ …

2. The cast

• N personas (one per axis), independent quality-maximizers: each proposes concrete candidates that push its quality as far as possible and judges everything from its axis.
• Synthesizer — one neutral voice above the panel. Not a quality-maximizer; a critique-guided integrator. It prunes, de-duplicates, recombines across axes, holds the frontier, and detects convergence. The only convergence authority (keeps the stop signal clean).

Frame-hardening (both modes). Personas and Synthesizer operate in a pure brainstorming frame. Tell each explicitly: ignore any repository, file, or coding-task context; you are a brainstorming voice, not a software agent. Without this, general-purpose subagents leak meta ("this is a self-contained task, no codebase work needed"). Keep their output to the PROPOSE/CRITIQUE/shortlist format only.

The isolation rule (isolated mode)

In isolated mode, personas are independent thinkers: they never see each other's chain of thought or system prompt — only the relayed candidate pool (conclusions, not reasoning). No live persistent agent across rounds: re-spawn a fresh agent each round and replay that persona's own prior turns into its prompt, so each accumulates memory of its own line without ever seeing how the others think. Relay candidates and critiques, never internal reasoning. (In batched mode this is enforced by instruction inside the single call instead.)

(Mechanism note: SendMessage live continuation is unavailable in this harness; re-spawn + own-history replay is equivalent here, since personas exchange only final candidates/critiques.)

Spawn with subagent_type: general-purpose. Calling the skill is the authorization to spawn.

3. Round loop (default 2 rounds, hard cap 3; early-stop on a stable frontier)

Maintain the current candidate pool (each candidate tagged with its provenance — which axis/axes authored it); in isolated mode also keep one running transcript per persona.

Per round:

1. Personas — one combined turn each (isolated: N spawns; batched: 1 spawn covering all). Each persona returns, in one turn:
   • PROPOSE: 1–3 new concrete candidates that maximize its quality, and
   • CRITIQUE: the single sharpest weakness in other candidates, judged from its axis. (Round 1 has nothing to critique — propose only. The critique is what does the real pruning, so a one-round run is the weak run; prefer 2 rounds when the decision matters.)
2. Synthesizer. Spawn with the full pool + this round's proposals/critiques. It:
   • prunes weak candidates — driven by the round's critiques (collisions, dead-on-an-axis), plus strict domination where it actually applies, and de-duplicates near-identical proposals from different personas,
   • recombines across axes — crossover: fuse the strengths of candidates from different axes into new hybrids (the memorable one's sound + the meaningful one's root), tagging the hybrid's provenance with all contributing axes — this is where the best answers come from,
   • updates the frontier, and emits a one-line digest.
3. Stop when the frontier is unchanged from the prior round (no new survivor), or at round 3.

Running digest

One tight line per round, emitted as you go so the user can watch and interrupt:

R1  frontier: «cand-a»⟨mem,own⟩ «cand-b»⟨mean⟩ «cand-c»⟨aes⟩   (+2 new, −3 pruned)
R2  frontier: …

4. Output — the short list (no voting)

When the frontier stabilizes or the cap is hit, the Synthesizer returns the short list: the surviving candidates, each annotated with which axes it maximizes and which it sacrifices, plus one line of recommendation:

SHORTLIST
  1. «candidate»  — wins: memorability, ownability   sacrifices: literal meaning
  2. «candidate»  — wins: meaning, pronounceability  sacrifices: distinctiveness
  3. «candidate»  — balanced; no axis loves it, none hates it

PICK (if forced): #1 — strongest on the two axes hardest to fix later; meaning can be supplied by use.

No election, no forced winner — the user decides. The one-line pick is advisory and ignorable. Return inline (these outputs are usually ephemeral; write a file only if the user asks).

Notes

• Cost is dominated by spawn count, not output length. Measured: ~27k tokens per spawn, flat, whether the reply is 40 words or 220. Isolated N=3 × 2 rounds ≈ 16 spawns ≈ 430k tokens; one isolated round ≈ 4 spawns ≈ 110k; batched ≈ 2 spawns/round. Default to batched; reserve isolated mode for genuinely consequential calls.
• Crossover is the highest-value step — most picked answers are hybrids the Synthesizer built, not raw persona proposals. Protect it.
• In isolated mode run personas in parallel within a round (they're independent); the Synthesizer runs after. Rounds are serial.
• If the question turns out to have a ground-truth answer, or needs research not judgment, say so and stop — this skill is for genuine no-right-answer choices.