Expand the option space before you decide.
🌐 English | 日本語 (Japanese)
A decision methodology distilled from the decision-workbench project: before making an important decision, use a heterogeneous multi-model committee to widen the option space, structure the divergence, then fold real constraints back in — and let a human make the final call.
- Decision power = size of the option space, not the ability to pick. Your cognitive surface is your decision ceiling.
- A single model (or one model in many roles) only re-samples its own training distribution. True divergence requires different model families (Anthropic / OpenAI / DeepSeek / Google / local OSS), isolated from each other to prevent anchoring.
- Model disagreement is often post-hoc rationalization + ecosystem bias — but ecosystem blind spots are exactly the highest-value divergence (e.g. CN models raise ICP/WeChat/Aliyun; Western models default to AWS).
- Feed the same brief to models from different families, isolated (invisible to each other).
- Assign stances per model: architecture / cost / security / product → natural divergence.
- Mandatory question patterns:
- "Why didn't you choose X?" — forces out excluded options + reasons (highest information density)
- "Give one non-mainstream option with a real justification." — kicks the model off its training distribution
- "Where is this plan most vulnerable?" — cross-examination a single model never performs on itself
- Consensus zone → adopt directly (standard-answer parts of the system).
- True divergence zone → worth human arbitration.
- Pseudo-divergence zone → noise, discard (e.g. PHP vs Go for a CRUD system is not a material difference).
- Fold back real constraints: budget, team skills, timeline, market, compliance.
- Optional:
pyDecision(AHP/TOPSIS/PROMETHEUS) turns options into comparable score tables. - Final review is always human. Models fail confidently.
- "AI shouldn't make important decisions for you. It should make the decision space harder for you to misunderstand."
- The committee expands your known unknowns; it cannot reach unknown unknowns — and being forced to understand the divergence while converging is the cognitive growth.
- Best posture: ask AI when you don't know what you don't know, not when you already know what you want.
- Don't treat the committee as an "auto-optimal-answer" machine — open tools (MoA / LLM-Blender) fuse answers by default and delete the human arbitration step, which is the opposite direction.
- Pseudo-divergence wastes time: models will fabricate complete rationales for any assigned stance.
- When requirements don't constrain the choice, models output their own familiar default — not an optimum.
- Models silently assume hidden constraints (concurrency, budget, team); humans must backfill the real ones.
This repository ships the methodology in the Agent Skills format (a single Markdown file with YAML frontmatter). It works with any tool that can consume instruction files — no code, no dependencies.
| Tool | Where to put it |
|---|---|
| Claude Code | ~/.claude/skills/decision-space-expansion/SKILL.md (global) or .claude/skills/decision-space-expansion/SKILL.md (per-project) |
| Cursor / Windsurf | .cursor/skills/decision-space-expansion/SKILL.md, or .cursor/rules/ |
| OpenAI Codex | Put the core method into AGENTS.md, or point Codex at the file as reference docs |
| Gemini CLI | ~/.gemini/skills/ (agent skills are supported) |
| Any agent / chat | Paste the SKILL.md content into the conversation, or cat SKILL.md and follow it |
Note:
SKILL.mdis tool-agnostic. Swap "different model families" for whatever pool your environment can reach (aggregator, multi-CLI, local Ollama/vLLM) — the method stays the same.
README.md— English summaryREADME.ja.md— 日本語サマリー (Japanese summary)SKILL.md— full skill definition (frontmatter + body), consumable by Claude Code, Cursor, Codex, Gemini CLI, Hermes, and any Markdown-reading agent
Related: the reference implementation lives in the decision-workbench repository (FastAPI + React, mock zero-key, 31 tests, hidden-gold eval: single-model 26% → naive multi-agent 78% → decision-workbench 100% / 12).
MIT