Google gives you sources but no synthesis — you drown in tabs. ChatGPT gives you synthesis but no sources — you can't verify anything, and it vanishes when you close the window.
This skill gives you both. One command, one markdown file where every claim links back to its source.
/deep-research Dario Amodei's vision for AI safety
You get a structured playbook you can keep, share, and update later.
Heads up: Each run launches 3-5 parallel Opus agents. Token usage is higher than you'd expect. Best suited for Claude Max subscribers.
/deep-research Harness Engineering for AI Agents
/deep-research Agent-Browser
/deep-research Dario Amodei
/deep-research React Server Components --file rsc.md
/deep-research --file existing-playbook.md
| Input | What happens |
|---|---|
| Just a topic | Creates <topic-slug>.md in your current directory |
Topic + --file path.md |
Writes to the specified path |
--file existing.md only |
Updates the existing playbook with new findings |
npx skills add webcpu/deep-research-skill -a claude-code --globalThen type /deep-research in Claude Code. That's it.
Or copy manually:
mkdir -p ~/.claude/skills/deep-research
curl -o ~/.claude/skills/deep-research/SKILL.md \
https://raw.githubusercontent.com/webcpu/deep-research-skill/main/deep-research/SKILL.md- Parse — detects whether you gave a topic, a file path, or both
- Scope — decides create vs. update mode; picks research angles
- Research — launches parallel agents with
run_in_background: true, each searching from a different angle - Synthesize — filters findings (drops anything without a source URL), groups by theme, writes the playbook
- Verify — reads the result to check for duplicates, missing citations, and cohesion
Every playbook follows this structure:
# Topic Title
## Section Name
Insight text. — [Source Name](https://url)
Another insight. — [Another Source](https://url)
---
## Sources
- [Source 1](https://url)
- [Source 2](https://url)
---
*Captured: 2026-03-03*Every claim is cited. No filler, no hedging.
/deep-research Apple Studio Display XDR produced a 14-section playbook with 36 sources and ~65 integrated findings:
- What It Is — Overview, pricing, availability
- Display Specifications — Full comparison table (XDR vs Studio Display vs Pro Display XDR)
- Configurations and Pricing — All SKUs, nano-texture analysis
- Connectivity — Thunderbolt 5, daisy-chaining, charging watts
- 120Hz Compatibility — Which Macs get 120Hz vs 60Hz
- Professional Reference Modes — 16 built-in presets, custom modes, DICOM medical imaging
- Camera, Audio, and Edge Light — Webcam history, speakers, virtual ring light
- The Pro Display XDR Trade-Off — Gains and losses vs the discontinued 32" model
- Known Issues and Gotchas — Stand lock-in, power cable, cable length, thermal throttling, firmware bricking, sleep/wake bugs
- Alternatives and Competitors — 6 alternatives with pricing
- Who Should Buy What — Decision framework by use case and Mac chip
- Calibration and Color Accuracy — Factory deltaE data, hardware calibration
- Compatibility Requirements — OS and hardware requirements
- Environmental — Recycled materials
See the full output: examples/apple-studio-display-xdr.md
Point it at a file you've already created:
/deep-research --file agent-browser.md WebDriver BiDi protocol
It reads the existing content, launches agents focused on the new topic, and merges findings into the existing sections — matching the voice and style already in the document. It never overwrites or creates a separate "Updates" section.
- Claude Code with Opus
- Claude Max recommended — each run launches 3-5 parallel Opus agents. Token usage is higher than you'd expect.
MIT