brains-trust

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.

Defaults (When User Just Says "Brains Trust")

If the user triggers this skill without specifying what to consult about, apply these defaults:

1. Pattern: Consensus (2 models from different providers) — it's called "brains trust", not "single opinion"
2. Scope: Whatever Claude has been working on in the current session. Look at recent context: files edited, decisions made, architecture discussed, problems being solved.
3. Mode: Infer from context:
   • Recently wrote/edited code → Code Review
   • In a planning or design discussion → Architecture
   • Debugging something → Debug
   • Building prompts or skills → Prompting
   • No clear signal → General (ask: "what are we missing? what are our blind spots?")
4. Models: Pick the newest pro-tier model from 2 different providers (check models.flared.au). Prefer diversity: e.g. one Google + one OpenAI, or one Qwen + one Google. Never two from the same provider.
5. Prompt focus: "Review what we've been working on. What are we missing? What could be improved? What blind spots might we have? Are there simpler approaches we haven't considered?"

Trigger → Default Mapping

Trigger	Default pattern	Default scope
"brains trust"	Consensus (2 models)	Current session work
"second opinion"	Single (1 model)	Current session work
"ask gemini" / "ask gpt"	Single (specified provider)	Current session work
"peer review"	Consensus (2 models)	Recently changed files
"challenge this" / "devil's advocate"	Devil's advocate (1 model)	Claude's current position

The user can always override by being specific: "brains trust this config file", "ask gemini about the auth approach", etc.

Setup

Set at least one API key as an environment variable:

# 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	Default for	What happens
Consensus	"brains trust", "peer review"	Ask 2 models from different providers in parallel, compare where they agree/disagree
Single	"second opinion", "ask gemini", "ask gpt"	Ask one model, synthesise with your own view
Devil's advocate	"challenge this", "devil's advocate"	Ask a model to explicitly argue against your current position

For consensus, always pick models from different providers (e.g. one Google + 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

1. Detect available keys — check OPENROUTER_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY in environment. If none found, show setup instructions and stop.

2. 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.

3. Read target files into context (if code-related). For non-code questions (strategy, prompting, general), skip file reading.

4. 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.

5. Create consultation directory at .jez/artifacts/brains-trust/{timestamp}-{topic}/ (e.g. 2026-03-10-1423-auth-architecture/). Write the prompt to prompt.txt inside it — never pass code inline via bash arguments (shell escaping breaks it).

6. Generate and run Python script at .jez/scripts/brains-trust.py using patterns from references/provider-api-patterns.md:

   • Reads prompt from the consultation directory's prompt.txt
   • Calls the selected API(s)
   • For consensus mode: calls multiple APIs in parallel using concurrent.futures
   • Saves each response to {model}.md in the consultation directory
   • Prints results to stdout

7. 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

1. Never hardcode model IDs — always fetch from models.flared.au first
2. Never cap output tokens — don't set max_tokens or maxOutputTokens
3. Always write prompts to file — never pass via bash arguments
4. Include file contents inline — attach code context directly in the prompt
5. Use AI-to-AI framing — the model is advising Claude, not talking to the human
6. Print progress to stderr — the Python script must print status updates (Calling gemini-2.5-pro..., Received response from qwen3.5-plus.) so the user knows it's working during the 30-90 second wait

Reference Files

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