Documentation

Overview

FastSkills is an open, searchable catalog of reusable skills for AI assistants — packaged instructions that teach a model how to perform a specific task well. Every skill is a plain Markdown file you can read, download, clone and version. The library is organised into four domains — Finance, Trading, Legal and Marketing — each with a second level of sub-labels for quick navigation.

Skills follow the open SKILL.md format, so a skill you find here works with any assistant that supports it — including Claude, ChatGPT and Grok.

What is a skill?

A skill is a self-contained set of instructions that gives an assistant specialised, repeatable expertise for one job — a legal review checklist, a deal-screening procedure, an SEO audit, a backtesting workflow. Instead of re-explaining the same process every time, you hand the model a skill and it follows a consistent, auditable procedure.

Concretely, a skill is a folder whose entry point is a SKILL.md file. That file carries a little metadata plus the instructions themselves. A skill can also bundle supporting files:

my-skill/
├── SKILL.md        # required: metadata + instructions
├── references/     # optional: background material
├── scripts/        # optional: helper code
└── assets/         # optional: templates, examples

On FastSkills, the SKILL.md body is the part you read on a skill's page and download; larger skills link out to their full source.

The SKILL.md format

A SKILL.md file starts with a YAML frontmatter block (between --- lines) and is followed by the instructions in Markdown. FastSkills reads these frontmatter fields:

A minimal example

---
title: NDA Review
description: Review a one-way NDA and produce a clause-by-clause issue log.
category: Legal
sublabel: Contracts
author: Your Name
tags: nda, contracts, review
license: MIT
---

# NDA Review

## When to use
- Reviewing a one-way commercial NDA before signature.

## Steps
1. Identify the parties, term, and definition of Confidential Information.
2. Flag one-sided or unusual clauses with a preferred redline.
3. Output an issue log: clause, risk, recommendation.

A good body answers three questions: when to use the skill, the steps to follow, and the output it should produce. Keep it specific and action-oriented — the clearer the procedure, the more reliably a model follows it.

How skills load

Skills are designed to stay out of the way until they're needed, through progressive disclosure:

• Discovery — an assistant first sees only each skill's name and description, enough to know when it might be relevant.
• Activation — when a task matches, it reads the full SKILL.md instructions into context.
• Execution — it follows the steps, loading any referenced files or running bundled scripts only if the task calls for them.

Because full instructions load only on demand, an assistant can keep many skills available while spending very little context on the ones it isn't using.

Categories & sub-labels

Every skill sits in one of four categories, each split into sub-labels so you can drill down quickly. Pick a category tab on the home page, then a sub-label chip to narrow further.

Sub-labels are assigned with a documented taxonomy and labelling prompt kept in the repository, so classification stays consistent as the library grows. When you author a skill, the editor suggests the sub-labels for its category.

Browsing & searching

The home page is the catalog. You can:

• Search titles, descriptions, tags and content from the search bar.
• Filter by category (the tabs) and then by sub-label (the chips).
• Open any card to read the full skill, or use the clone button to fork it.

Filters are just URL parameters, so links are shareable — for example /?category=Finance&sub=Family%20Office or /?q=backtest.

Using a skill

Open a skill and choose Download SKILL.md to get the file. From there it works with any assistant that reads the format:

• Claude / Claude Code — place the skill in your skills directory (or a project), and Claude loads it when a task matches its description.
• ChatGPT — add the instructions to a Project or a Custom GPT so they apply to that workspace.
• Grok — paste the instructions into your custom/system instructions.

However you load it, the pattern is the same: the metadata tells the assistant when the skill is relevant, and the body tells it exactly what to do. Nothing on FastSkills is provider-specific — the same skill is portable across tools.

Cloning a skill

Every skill — on its detail page and on each catalog card — has a ⑂ Clone button, like a fork on a code host. Cloning:

• Creates a private draft that you own, pre-populated with the original's title, content, category, sub-label and tags.
• Opens it straight in the editor so you can adapt it.
• Records where it came from — the clone shows a “forked from” link back to the original, preserving attribution.

If you're signed out, the clone button opens the sign-in dialog first; once you're in, clone again and your copy is created.

Versioning

FastSkills keeps two kinds of version:

• Default (library) versions live in GitHub. The catalog's built-in skills are seeded from the repository, so their canonical history is the commit history — library skills carry a badge and a “View in GitHub” link.
• Your versions live in the database. Every time you save a skill you own, a new version is recorded.

Open History on any skill to see the full list of versions with timestamps. You can view any earlier version read-only, and — on skills you own — restore one, which brings it back as the current version while keeping the prior state in the history.

Creating & editing

Use New Skill to start from scratch, or clone an existing one. The editor gives you three interchangeable writing surfaces for the body:

• Rich — a formatted editor for prose, lists, tables and code.
• Blocks — a block-by-block view for restructuring.
• Markdown — raw Markdown, if you prefer to write it directly.

Alongside the body you set the metadata — category, sub-label (with suggestions), author label, tags, a short description and visibility. Changes autosave as you type, and each save creates a version. When it's ready, Publish it; you can always move it back to draft.

Public & private

Every skill has an owner and a visibility setting:

• Public (the default) — listed in the catalog for everyone once published.
• Private — visible only to you; useful for drafts or internal skills.

Your skills — public and private, drafts included — live under My Skills. Author attribution (the name on the card) is separate from ownership, so you can credit an original author while owning your own copy.

Contributing

There are two ways to add to FastSkills:

• Create a skill in the app and publish it public — it appears in the catalog immediately, owned by you.
• Contribute a library skill via the repository: add a SKILL.md under the seed/ tree and open a pull request. Library skills are re-applied on deploy, so an approved change lands automatically.

The catalog and the repository stay in sync: library skills are the default versions in GitHub; anyone can clone them into their own editable copy.

API reference

A read-only JSON API is available for the public catalog:

Interactive docs live at /api/docs.

Self-hosting

FastSkills is open source. It's a single FastHTML application backed by SQLite or PostgreSQL, packaged as a small Docker image. You can run your own private catalog for a team, seeded from your own seed/ tree.

See the repository for setup, configuration and deployment instructions.

FAQ

Do I need an account to browse?

No — browsing, searching and downloading are open. You need an account only to clone, create or edit skills.

Which assistants does a skill work with?

Any that support the SKILL.md format, including Claude, ChatGPT and Grok. The content is provider-neutral.

Who can edit a library skill?

Library (default) skills are maintained in GitHub. To make your own changes, clone one — your edits become a versioned copy you own in the database.

How are sub-labels decided?

From a fixed per-category taxonomy and a labelling prompt kept in the repository, so classification is consistent and reproducible.