skill-brainstorm

Two AI models argue so you don't have to. One idea in, one decision out.

Claude runs the show. Gemini pushes back. Facts get checked. You pick the winner.

---

The Problem

You have an idea and need to make a decision. You ask one AI -- it agrees with you. You ask another -- it also agrees. Everyone's polite, nobody challenges anything, and you ship something based on vibes.

This skill fixes that. It runs a structured 3-round adversarial dialogue between Claude and Gemini. Ideas get challenged, weak ones get killed, facts get verified. You end up with one concrete action -- not a list of "options to consider."

---

How It Works (Plain English)

Think of it as a board meeting between two opinionated experts:

Step	What Happens	You See
Context	You describe what you're building and your constraints	Claude asks clarifying questions if needed
Research	Both Claude and Gemini search the web in parallel to verify current facts	A "Verified Context" block with real data
Round 1: Diverge	Gemini challenges your framing and proposes 5-7 alternative angles	New ideas you hadn't considered
Round 2: Deepen	Claude kills the weak ideas with evidence. Gemini defends or concedes	The field narrows from 7 to 2-3 survivors
Check-in	You're asked: "Which direction do you prefer?"	Your input before the final round
Round 3: Converge	One winner is chosen with a timeline and success criteria	A single action, not a strategy deck
Fact-check	All claims in the final decision are verified against the live web	A confidence table (confirmed/conflict/incorrect)

The whole process takes 3-5 minutes and costs about $0.35 in API calls.

---

Install

Claude Code (recommended):

/install github:awrshift/skill-brainstorm

Manual:

mkdir -p .claude/skills/brainstorm/scripts
curl -sL https://raw.githubusercontent.com/awrshift/skill-brainstorm/main/SKILL.md \
  -o .claude/skills/brainstorm/SKILL.md
curl -sL https://raw.githubusercontent.com/awrshift/skill-brainstorm/main/scripts/gemini.py \
  -o .claude/skills/brainstorm/scripts/gemini.py

Requirements

What	Why	How to Get
Google API Key	Powers the Gemini calls	Free at aistudio.google.com
google-genai package	Python SDK for Gemini	pip install google-genai
Claude Code	Orchestrates the brainstorm	You're probably already using it

Add your key to a .env file in your project:

GOOGLE_API_KEY=your-key-here

---

When to Use

Just say any of these to Claude:

• "Brainstorm how to launch X"
• "Let's think through options for Y"
• "I need to decide between A and B"
• "Explore this idea with Gemini"
• "Diverge and converge on this"

The skill triggers automatically. No special commands needed.

---

Architecture

The skill uses a two-layer model split -- a cheaper model handles research, an expensive model handles reasoning:

              Claude (orchestrator)
                 |         |
    Research     |         |     Reasoning
    phases       |         |     rounds
                 v         v
          Flash-Lite        Pro
       (web search,      (critical thinking,
        fact-check)       argumentation)
                 |              ^
                 +--Verified----+
                    Context

Why two layers? When one model does both searching AND thinking, it wastes expensive reasoning capacity on simple lookups like "what version is React?" Splitting the work means:

• Flash-Lite -- searches the web, verifies facts, checks claims
• Pro -- all brainpower goes to challenging ideas, finding blind spots, building arguments

---

What You Get at the End

Every brainstorm produces:

1. One concrete action -- not 5 options, one thing to do next
2. A timeline -- hours or days, not vague "weeks"
3. Go/no-go criteria -- "if X doesn't happen by Y, abandon this"
4. Kill list -- everything that was considered and why it was rejected
5. Fact-check table -- which claims were verified, which had conflicts
6. The full journey -- 7+ ideas narrowed to 3, then to 1

---

Gemini API Cost

Each brainstorm makes 5-6 Gemini API calls. Actual cost depends on how long your prompts and responses are. Here are the per-million-token rates so you can estimate:

Model	Role in Brainstorm	Calls	Input / 1M tokens	Output / 1M tokens
Flash-Lite (gemini-3.1-flash-lite-preview)	Research + fact-check	2-3	$0.25	$1.50
Pro (gemini-3.1-pro-preview)	3 reasoning rounds	3	$2.00	$12.00

Claude orchestration is included in your Claude Code session -- no extra cost.

A typical brainstorm uses roughly 5-15K tokens per Gemini call. Pricing source: ai.google.dev/gemini-api/docs/pricing

---

Example Output

After a brainstorm about choosing a payment provider, you'd see something like:

Winner: Stripe Payment Links (not full Checkout integration)

Why it won: Zero backend code needed. Embed in landing page today. Upgrade to full Checkout only if conversion > 2%. Lemon Squeezy killed in R2 (no webhook reliability data). Paddle killed in R1 (EU-only tax handling, you need global).

Timeline: 2 hours to first payment link. 1 day to embed + test.

Go/no-go: If < 10 payments in first week, switch to direct Stripe Checkout.

Fact-check: 5/5 claims confirmed (Stripe pricing, Paddle EU limitation, Lemon Squeezy webhook issues).

---

Gotchas

• Requires GOOGLE_API_KEY -- without it, Gemini calls fail silently
• Package is google-genai not google-generativeai -- the SDK was renamed
• User check-in after Round 2 is mandatory -- prevents the models from converging on something you don't care about
• Works for any decision -- not just technical. Hiring, marketing, product strategy, architecture

---

Part of the AWRSHIFT Ecosystem

Skill	What It Does
skill-brainstorm	You're here. 3-round adversarial brainstorm
skill-gemini	Gemini inside Claude Code -- ask, ground, image, review
skill-awrshift	Adaptive decision framework (Quick / Standard / Scientific)
skill-telegram	Telegram: read, parse channels, voice notes
claude-starter-kit	All skills pre-configured + project template

---

Contributing

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

License

MIT -- see LICENSE for details.

---

One idea in. One decision out.