brains-trust

Original：🇺🇸 English

Translated

Get a second opinion from leading AI models on code, architecture, strategy, prompting, or anything. Queries models via OpenRouter, Gemini, or OpenAI APIs. Supports single opinion, multi-model consensus, and devil's advocate patterns. Trigger with 'brains trust', 'second opinion', 'ask gemini', 'ask gpt', 'peer review', 'consult', 'challenge this', or 'devil's advocate'.

25installs

Sourcejezweb/claude-skills

Added on2026-03-11

NPX Install

npx skill4agent add jezweb/claude-skills brains-trust

SKILL.md Content

View Translation Comparison →

Brains Trust

Consult other leading AI models for a second opinion. Not limited to code — works for architecture, strategy, prompting, debugging, writing, or any question where a fresh perspective helps.

Setup

Set at least one API key as an environment variable:

bash

# Recommended — one key covers all providers
export OPENROUTER_API_KEY="your-key"

# Optional — direct access (often faster/cheaper)
export GEMINI_API_KEY="your-key"
export OPENAI_API_KEY="your-key"

OpenRouter is the universal path — one key gives access to Gemini, GPT, Qwen, DeepSeek, Llama, Mistral, and more.

Current Models

Do not use hardcoded model IDs. Before every consultation, fetch the current leading models:

https://models.flared.au/llms.txt

This is a live-updated, curated list of ~40 leading models from 11 providers, filtered from OpenRouter's full catalogue. Use it to pick the right model for the task.

For programmatic use in the generated Python script:

https://models.flared.au/json

Consultation Patterns

Pattern	When	What happens
Single (default)	Quick second opinion	Ask one model, synthesise with your own view
Consensus	Important decision, want confidence	Ask 2-3 diverse models in parallel, compare where they agree/disagree
Devil's advocate	Challenge an assumption	Ask a model to explicitly argue against your current position

For consensus, pick models from different providers (e.g. one Google, one OpenAI, one Qwen) for maximum diversity of perspective.

Modes

Mode	When	Model tier
Code Review	Review files for bugs, patterns, security	Flash
Architecture	Design decisions, trade-offs	Pro
Debug	Stuck after 2+ failed attempts	Flash
Security	Vulnerability scan	Pro
Strategy	Business, product, approach decisions	Pro
Prompting	Improve prompts, system prompts, KB files	Flash
General	Any question, brainstorm, challenge	Flash

Pro tier: The most capable model from the chosen provider (e.g.

google/gemini-3.1-pro-preview

,

openai/gpt-5.4

). Flash tier: Fast, cheaper models for straightforward analysis (e.g.

google/gemini-3-flash-preview

,

qwen/qwen3.5-flash-02-23

).

Workflow

Detect available keys — check
```
OPENROUTER_API_KEY
```
,
```
GEMINI_API_KEY
```
,
```
OPENAI_API_KEY
```
in environment. If none found, show setup instructions and stop.
Fetch current models —
```
WebFetch https://models.flared.au/llms.txt
```
and pick appropriate models based on mode (pro vs flash) and consultation pattern (single vs consensus). If user requested a specific provider ("ask gemini"), use that.
Read target files into context (if code-related). For non-code questions (strategy, prompting, general), skip file reading.
Build prompt using the AI-to-AI template from references/prompt-templates.md. Include file contents inline with
```
--- filename ---
```
separators. Do not set output token limits — let models reason fully.
Write prompt to file at
```
.claude/artifacts/brains-trust-prompt.txt
```
— never pass code inline via bash arguments (shell escaping breaks it).
Generate and run Python script at
```
.claude/scripts/brains-trust.py
```
using patterns from references/provider-api-patterns.md:
- Reads prompt from
```
.claude/artifacts/brains-trust-prompt.txt
```
- Calls the selected API(s)
- For consensus mode: calls multiple APIs in parallel using
```
concurrent.futures
```
- Saves each response to
```
.claude/artifacts/brains-trust-{model}.md
```
- Prints results to stdout
Synthesise — read the responses, present findings to the user. Note where models agree and disagree. Add your own perspective (agree/disagree with reasoning). Let the user decide what to act on.

When to Use

Good use cases:

Before committing major architectural changes
When stuck debugging after multiple attempts
Architecture decisions with multiple valid options
Reviewing security-sensitive code
Challenging your own assumptions on strategy or approach
Improving system prompts or KB files
Any time you want a fresh perspective

Avoid using for:

Simple syntax checks (Claude handles these)
Every single edit (too slow, costs money)
Questions with obvious, well-known answers

Critical Rules

Never hardcode model IDs — always fetch from
```
models.flared.au
```
first
Never cap output tokens — don't set
```
max_tokens
```
or
```
maxOutputTokens
```
Always write prompts to file — never pass via bash arguments
Include file contents inline — attach code context directly in the prompt
Use AI-to-AI framing — the model is advising Claude, not talking to the human

Reference Files

When	Read
Building prompts for any mode	references/prompt-templates.md
Generating the Python API call script	references/provider-api-patterns.md