Agent Skills
AnthropicThe reference collection: document, spreadsheet, and slide skills from the team that defined the format.
Every entry is checked against the vendor's own documentation, dated, and given a one-line verdict on who it actually suits. No scraped blurbs, no dead links, no "top 100" padding.
A set of product-specific servers from Cloudflare covering observability, DNS, Workers, and account configuration. They are hosted, so there is no local process to k…
Searches and reads Confluence pages and spaces through the API. The value is retrieval: the answer to a process question usually exists in a page nobody searches by …
An official server exposing ElevenLabs text-to-speech, voice listing, and audio generation as tools. The useful pattern is narration inside a content pipeline: scrip…
Gives an agent scoped read, write, and search access to directories you explicitly allow. It is the reference implementation for local file work: you pass one or mor…
Searches and reads Drive files through a scoped OAuth grant: names, types, and extracted text for documents. It is the retrieval layer for a company's documents, whi…
The reference collection: document, spreadsheet, and slide skills from the team that defined the format.
Official skill for generating art as code — procedural patterns instead of image prompts.
Derive contract tests from an API description, not from the implementation.

The config file, the three values that actually matter, and how to know it worked.

A folder with a SKILL.md the agent loads on demand — closer to a training manual than a plugin.

The difference between a chatbot, a tool, and an agent — and why the distinction decides what you can delegate.

Most search is page retrieval. A different class of tool retrieves evidence — papers, filings, and publishers — and that changes what the answer…

Pi, ChatGPT, and Claude Code are all 'assistants'. They support three different modes of work, and picking the wrong mode is the whole problem.…

Relay and Lindy both automate multi-step work. The difference that matters is whether a person confirms before the irreversible step. Automation fails in two…
Lindy runs server-side agents over email and calendars; Bardeen automates the browser tab you are looking at.
Relay has approval steps built in; Zapier has the connectors. The risk shape decides.
Motion plans your time; Superhuman makes your inbox fast. Two different bottlenecks.
Every entry is dated. Reviewed entries were read by a person; the rest are compiled from the npm registry and labelled. How verification works →