Think Tank

CLI multi-model deliberation tool. Sends one prompt to Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and DeepSeek V4 Pro in parallel through OpenRouter. Optional crux framing, focused deliberation rounds, structured analysis, and Grok 4.20 chairman synthesis. Every run saves a markdown transcript and a polished HTML report by default.

Setup

• API Key: Set OPENROUTER_API_KEY in ~/.env, ~/.think_tank.env, or ./.env (searches multiple locations; also works with Windows user environment variables)
• Shell alias: Add to your shell profile so think_tank works from any directory:

  # Bash (~/.bashrc)
  alias think_tank="python3 /path/to/think-tank/think_tank.py"
  
  # Zsh (~/.zshrc)
  alias think_tank="python3 /path/to/think-tank/think_tank.py"
  
  # PowerShell ($PROFILE)
  function think_tank { python "C:\path\to\think-tank\think_tank.py" @args }

• Context: Auto-detects CLAUDE.md when run from inside a repo directory

Usage

think_tank "Your question here"

Run from inside a repo directory to auto-include project context.

Pipeline

Stage 0: CRUX        Optional framing pass (--crux)
Stage 1: COLLECT      Default council answers in parallel
Stage 2: DELIBERATE   Optional focused refinement rounds (--rounds)
Stage 3: ANALYZE      Chairman extracts consensus, disagreements, unresolved gaps
Optional: REVIEW      Leave-one-out anonymized peer ranking diagnostics (--review)
Stage 4: SYNTHESIZE   Grok 4.20 chairman compiles the final answer
Stage 5: SAVE         Markdown transcript + readable HTML report

Synthesis runs by default. Use --no-chairman to skip it (analysis still runs if 2+ models responded). Peer review/ranking is off by default; use --review only when you explicitly want model-grading diagnostics.

Flags

Flag	Short	Value	What it does
--chairman	-c	model key	Which model synthesizes the final answer (default: grok)
--blind	-b	—	Hide model identities until reveal at end
--no-chairman		—	Skip chairman synthesis; analysis still runs with 2+ responses
--review		—	Run optional anonymized peer review/ranking diagnostics
--crux		—	Run a framing pass before collection
--red-team MODEL	-R	model key	Assign one model an adversarial critique role
--deep	-d	—	Include MEMORY.md from .claude project memory
--rounds N	-r N	number	Deliberation rounds after the initial answer pass (default: 0)
--files x,y	-f x,y	paths	Include specific files as context
--save [path]	-s [path]	optional filepath	Choose the markdown/HTML artifact basename. Bare filenames go under output/; omit path for a timestamped file.
--no-save		—	Skip markdown/HTML artifact writes for a throwaway run.
--interactive	-i	—	Open-ended mode with follow-up prompts
--models a,b	-m a,b	names	Use specific models only (claude, gpt, gemini, deepseek, grok)
--no-context	—	—	Skip auto project context; explicit --files still load
--prompt-file	-pf	filepath	Read question from a file instead of command line
--json	—	—	Output structured JSON, suppress terminal display

Models

Key	Model	Chairman-eligible
claude	Claude Opus 4.7	Yes
gpt	GPT-5.5	Yes
gemini	Gemini 3.1 Pro	Yes
deepseek	DeepSeek V4 Pro	Yes
grok	Grok 4.20	Yes (default chairman, not in the default panel)

Examples

# Full pipeline — default 4-model panel, analysis, Grok chairman synthesis
think_tank "Should we split this into microservices?"

# Include specific source files for a code question
think_tank -f src/App.jsx,src/utils.js "How should we refactor the tab system?"

# Crux framing before collection
think_tank --crux "What's the right caching strategy here?"

# One focused deliberation round before synthesis
think_tank -r 1 "What's the right caching strategy here?"

# Optional model-grading diagnostics when you explicitly want them
think_tank --review "Which answer is strongest?"

# Adversarial critique from one model
think_tank -R gpt "Where could this migration plan fail?"

# Blind mode — hide model identities from yourself
think_tank -b "Which database should we use for this workload?"

# Use Claude as chairman instead of the default Grok
think_tank -c claude "Compare these two architectures"

# Skip synthesis
think_tank --no-chairman "Quick sanity check"

# Deep mode with crux framing and transcript saved
think_tank -d --crux -r 1 -s architecture.md "Long-term scaling plan?"

# Save to an explicit directory instead of output/
think_tank -s ~/transcripts/architecture.md "Long-term scaling plan?"

# Throwaway run with no markdown/HTML souvenirs
think_tank --no-save "Quick gut check?"

# Interactive session
think_tank -i "Let's design a new feature"

# General question (no project context)
think_tank --no-context "Compare Redis vs Memcached for session storage"

# Only ask Claude and DeepSeek
think_tank -m claude,deepseek "Quick sanity check on this approach"

API Key

This repo now uses OpenRouter only. Set:

• OPENROUTER_API_KEY

Direct provider keys are no longer used by think_tank.py.

How It Works

1. Auto-detects CLAUDE.md by walking up from current directory unless --no-context is set.
2. Loads explicit --files even when auto context is disabled.
3. If --crux is set, extracts cruxes, assumptions, and validation tests before collection.
4. Sends your question + context to all models in parallel.
5. If --rounds N is set, runs N deliberation rounds after the first answer pass.
6. Analysis: Chairman extracts consensus points, disagreements, unresolved gaps, and crux coverage.
7. Optional Review: With --review, anonymizes responses as Response A/B/C/D and runs leave-one-out peer ranking diagnostics.
8. Synthesis: Chairman compiles the final answer from the responses and structured analysis. Optional review rankings are treated only as a weak diagnostic.
9. Saves a markdown transcript plus a standalone HTML report with the synthesis promoted to the top.
10. If --blind: reveals model identity mapping at the very end.

Tips

• Default: Four independent perspectives, structured analysis, synthesis, and saved artifacts.
• --crux: Best first add-on for architecture or strategy decisions.
• --rounds 1: One focused response-to-response refinement pass before synthesis.
• --rounds 2+: Usually overkill; use it only for deliberate stress-testing because later rounds tend to become meta-heavy.
• --review: Optional model-grading mode. Useful for diagnostics, noisy for decision memos.
• --no-chairman: When you want responses plus structured analysis without synthesis.
• --blind: Forces you to evaluate responses without model bias before the reveal
• --no-context: Skip CLAUDE.md/deep memory; use with --files when you want only explicit context
• cd into the repo first — context auto-detection needs to find CLAUDE.md
• Markdown and HTML artifacts save by default under output/; use --save name.md to choose the basename or --no-save for a disposable run.
• The saved artifacts include everything: responses, analysis, synthesis, and optional review diagnostics. The HTML report promotes the executive summary/final synthesis first and collapses the source material below.