From taco
Socratic brainstorming: explore ideas through questions, multi-model debate, and structured output
How this skill is triggered — by the user, by Claude, or both
Slash command
/taco:brainstormThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
```
/brainstorm <feature-name>
Ask the user 3-5 Socratic questions to understand the core intent:
Use AskUserQuestion tool for this — present focused questions with concrete options where possible.
Spawn an explorer agent to scan the codebase:
Spawn a critic agent to challenge the idea:
Spawn a researcher agent if the idea involves:
Gather diverse perspectives on the idea using 3 models in parallel.
Reuses call-model.sh (call_codex / call_gemini) and synthesize-reviews.sh patterns.
Claude (critic results): Reuse the risk analysis from Phase 3 as Claude's perspective.
Codex (call_codex): Implementation-focused prompt —
"Given this idea: {Problem + Solution summary}. Evaluate: (1) implementation complexity, (2) technical feasibility, (3) alternative approaches. Be specific and concise."
Gemini (call_gemini): Broad-perspective prompt —
"Given this idea: {Problem + Solution summary}. Provide: (1) similar solutions in other ecosystems, (2) perspectives or angles not yet considered, (3) potential blind spots. Be specific and concise."
Synthesis: Combine the 3 perspectives using synthesize-reviews.sh → save to .dev/specs/{feature-name}/debate.md
Consensus detection: Automatically determined by synthesize-reviews.sh:
3-model — all three models contributed2-model — two models contributedclaude-only — only Claude (Codex/Gemini unavailable)Graceful degradation: If Codex or Gemini CLI is unavailable, proceed with available models. If only Claude is available, skip debate file and carry Phase 3 results directly into Synthesis.
Combine all inputs into .dev/specs/{feature-name}/idea.md:
---
title: {Feature Name}
date: {YYYY-MM-DD}
status: brainstorm
---
## Problem
{What problem this solves, for whom}
## Solution
{Proposed approach, key ideas}
## Context
{Relevant existing code, patterns, constraints discovered by explorer}
## Risks
{Risks and concerns identified by critic}
## Debate Summary
{3-model debate synthesis. Points of agreement, disagreement, and key insights}
- **Consensus**: {what models agreed on}
- **Divergent**: {where models disagreed}
- **Key Insight**: {most valuable finding from debate}
## References
{External resources found by researcher, if any}
## Open Questions
{Unresolved questions for specify phase}
If debate was claude-only (no debate.md), write "Single-model analysis (Codex/Gemini unavailable)" in Debate Summary and summarize key points from the critic analysis.
After idea.md is written, perform a final review before handing off to /specify:
status: brainstorm → status: clarified..dev/specs/{feature-name}/call-model.sh / synthesize-reviews.sh — this is the only permitted Bash usagenpx claudepluginhub mildsalmon/taco-harness --plugin tacoGuides phased collaborative brainstorming: understand ideas via one-at-a-time questions, explore 2-3 approaches with trade-offs, validate incremental designs before implementation.
Guides structured brainstorming to clarify user intent, explore approaches, trade-offs, and refine requirements before implementing features or changes. Activates on ambiguous requests.
Guides collaborative exploration of feature requirements and design approaches before implementation planning.