home / skills / phrazzld / claude-config / thinktank

thinktank skill

safe

/skills/thinktank

This skill coordinates multi-model expert reviews to synthesize diverse perspectives into actionable recommendations for architecture, code reviews, and

npx playbooks add skill phrazzld/claude-config --skill thinktank

Review the files below or copy the command above to add this skill to your agents.

Files (1)

SKILL.md

1.1 KB

---
name: thinktank
description: |
  Multi-model expert review with synthesis.
  Runs multiple AI models in parallel for diverse perspectives.
  Use when: architecture decisions, code review, security audit, need consensus.
argument-hint: '"<query>" [file...]'
effort: high
---

# /thinktank

Multiple expert perspectives on any question.

## Role

Orchestrator gathering multi-model consensus.

## Objective

Answer `$ARGUMENTS` with diverse AI perspectives, synthesized into actionable recommendations.

## Workflow

1. **Frame** — Write clear instructions to temp file
2. **Context** — Include specified files or branch diff
3. **Run** — `thinktank /tmp/thinktank-query.md $FILES --synthesis`
4. **Synthesize** — Report consensus, divergent views, recommendations

## Usage

```
/thinktank "Is this auth implementation secure?" ./src/auth
/thinktank "What are the tradeoffs of this architecture?"
/thinktank "Review this PR for issues" $(git diff main --name-only)
```

## Output

- **Consensus** — What all models agree on
- **Divergent** — Where models disagree (investigate further)
- **Recommendations** — Prioritized actions

Overview

This skill runs multiple AI models in parallel to produce expert reviews and a synthesized consensus. It gathers diverse perspectives, highlights disagreements, and delivers prioritized, actionable recommendations for decisions and reviews. Use it to reduce blind spots and accelerate decision making with multi-model evidence.

How this skill works

The orchestrator frames the query into a temporary instruction file, includes relevant context such as files or diffs, and runs multiple models in parallel. Results from each model are collected, compared to identify consensus and divergence, and then synthesized into a concise report with prioritized recommendations. The final output separates what all models agree on, where opinions differ, and specific next steps.

When to use it

Architecture decisions where trade-offs must be weighed from multiple viewpoints
Code reviews or pull request audits that need broad coverage and prioritized fixes
Security audits to surface potential vulnerabilities and recommended mitigations
Complex design choices requiring evidence-based consensus and documented disagreements
Situations where you need a synthesized recommendation from several AI perspectives

Best practices

Provide focused, explicit prompts and include only necessary files or diffs to reduce noise
Frame the desired outcome (e.g., security hardening, performance improvement) to guide synthesis
Use the divergent section to assign follow-up tasks or deeper manual investigation
Prioritize recommendations into quick wins and longer-term actions for clear execution
Iterate: refine prompts and re-run with updated context after applying high-priority fixes

Example use cases

Ask for a security review of an authentication module and get consensus plus prioritized mitigations
Request a systems architecture trade-off analysis comparing scalability, cost, and complexity
Perform a multi-model code review on a PR to catch bugs, style issues, and design concerns
Audit dependency updates and receive synthesized guidance on compatibility and risk
Evaluate incident postmortems to surface common root causes and recommended process changes

FAQ

How are disagreements handled?

Disagreements are listed in a Divergent section with specifics on the conflicting points and suggested follow-up steps for investigation.

What output format should I expect?

You receive a report with three parts: Consensus (what all models agree on), Divergent (where they differ), and Recommendations (prioritized actions).