ChatGPT Work Tool and Skill Reference
TL;DR Highlight
A site publishes the full text of ChatGPT Work's 232 tools and 44 skills.
Who Should Read
An agent developer who wants a reference for writing tool descriptions and SKILL.md files
Core Mechanics
- 232 tools and 44 skills are published
- The snapshot is dated 31 Aug 2026.
- The skill source totals 615k characters.
- Tools and skills play different roles
- Tools are callable endpoints, like functions.
- Skills are reusable instruction packages on how to use those tools.
- Tools are grouped by category
- GitHub has 89, Sites 23, Gmail 21, Google Calendar 15.
- Runtime & files has 8 and Sub-agents & coordination has 6.
- Skill pages reproduce the full SKILL.md
- Tool pages keep the exposed descriptions and TypeScript declarations.
- Skill descriptions say when to use and when not to
- imagegen says not to use it for visuals better built as SVG or code.
- control-browser says to search first for public lookups and not fall back to Browser if search fails.
Community Reactions
- The control-browser skill structure stood out to a commenter
- It launches Playwright via a Node.js REPL, then calls browser.documentation() for further instructions.
- They wonder why the instructions aren't all in the SKILL.md.
- A commenter notes AI-generated sites all look alike
- They compare it to the Bootstrap era and ask if there is a shared guideline.
- A commenter says too many tools slow things down and waste tokens
- Some Work tools are said to add latency and token cost.
- A commenter hit a UI problem on the site
- The left sidebar doesn't scroll on its own, so on non-4k monitors the bottom of the list is unreachable.
How to Apply
- Add use-when / don't-use-when to your skill descriptions
- Open the SKILL.md or tool description you use most today.
- Add separate Use when and Do not use when lines, like imagegen.
- Test loading long instructions on demand
- Keep the core in SKILL.md and move details to a separate doc fetched at call time.
- Compare token usage with the documentation() approach.
- Group tools by category to limit exposure
- Split the tools you pass to the agent into per-task bundles.
- Enable only the needed bundle to cut token waste.
Try It
# Paste into Claude or ChatGPT (simplified version of the source's skill-description style)
Below is a tool/skill description from my agent.
{{tool or skill description}}
Rewrite it in this format.
1. One-line summary
2. Use when: 3 situations
3. Do not use when: 3 situations
4. Where it could be confused with similar tools
Next step: give it 5 ambiguous requests and ask which tool it would pick with the rewritten description.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.