claude-feynman

Research agent skills for Claude Code — adapted from Feynman.

Gives Claude Code a full research workflow: deep investigations, literature reviews, peer review simulations, paper-vs-code audits, experiment replication, and more. All orchestrated through slash commands using Claude Code's native tools (WebSearch, WebFetch, Agent subagents).

Install

Into the current project (recommended):

curl -fsSL https://raw.githubusercontent.com/shivchander/claude-feynman/main/install.sh | bash

Or manually:

git clone https://github.com/shivchander/claude-feynman.git /tmp/claude-feynman
cd /tmp/claude-feynman && bash install.sh /path/to/your/project

This installs:

• .claude/commands/ — 15 slash commands
• .claude/CLAUDE.md — research agent project config (appended if it already exists)
• Output directories — outputs/, papers/, notes/, experiments/

Commands

Research Workflows

Command	What it does
/deepresearch <topic>	Source-heavy multi-agent investigation with parallel researchers, synthesis, citation, and verification
/lit <topic>	Literature review from paper search and primary-source synthesis
/review <artifact>	Simulated peer review with FATAL/MAJOR/MINOR severity and inline annotations
/audit <paper>	Paper claims vs. codebase mismatch audit
/compare <topic>	Source comparison matrix with evidence grading
/draft <topic>	Paper-style draft from research findings

Experiments

Command	What it does
/replicate <paper>	Plan or execute experiment replication with environment selection
/autoresearch <idea>	Autonomous optimization loop: edit, measure, keep/revert, repeat

Utilities

Command	What it does
/eli5 <topic>	Plain-English explanation of a paper or concept
/log	Write a session log with findings, open questions, next steps
/watch <topic>	Set up recurring research monitoring
/jobs	Inspect active background work (tmux, processes, in-progress artifacts)

Compute

Command	What it does
/docker	Run experiments in isolated Docker containers
/modal	Serverless GPU compute on Modal
/runpod	Persistent GPU pods on RunPod with SSH access

How It Works

Each command is a Markdown prompt file in .claude/commands/. When you type /deepresearch scaling laws in Claude Code, it loads the deepresearch.md prompt with your topic and executes the full workflow.

The research workflows use Claude Code's Agent tool to spawn parallel research subagents — each gathering evidence independently, writing to files, and returning lightweight summaries. The lead agent synthesizes, writes, cites, verifies, and delivers.

Subagent roles:

• Researcher — gathers primary evidence across papers, web, repos, docs
• Writer — turns research notes into structured drafts
• Verifier — adds inline citations, verifies source URLs, removes unsourced claims
• Reviewer — simulated peer review with severity-graded feedback

Output structure:

outputs/            # Research briefs, comparisons, reviews
outputs/.plans/     # Workflow plans and task ledgers
outputs/.drafts/    # Intermediate drafts before citation
papers/             # Paper-style drafts and writeups
notes/              # Session logs and scratch notes
experiments/        # Experiment code and results

Every artifact follows a slug-based naming convention (<slug>.md, <slug>.provenance.md) and deep research outputs include provenance sidecars tracking sources consulted, accepted, and rejected.

Uninstall

rm -rf .claude/commands/{deepresearch,lit,review,audit,compare,draft,replicate,autoresearch,eli5,log,watch,jobs,docker,modal,runpod}.md

The CLAUDE.md config section is marked with comments — remove the block between # --- claude-feynman start --- and # --- claude-feynman end ---.

Credits

Adapted from Feynman by Companion AI. Research methodology, subagent patterns, and workflow designs originate from the Feynman project. Rewritten for Claude Code's native tool ecosystem.

License

MIT