The Trusted Open Library
of AI Skills

Size is easy. Trust is the point.

Abstract

A skill is a plain-text method (a single .md file) you hand to your AI assistant so it performs a professional task with a proven approach. A skill is also a set of instructions your agent will execute with your permissions — which makes every skill library a supply chain. This document describes a library built for that reality.

1.Problem statement

The scale is measured, not anecdotal. Scanning 3,984 skills across two public hubs on 5 February 2026, Snyk found that 36.82% carried at least one security flaw and 13.4% at least one critical issue — plus 76 payloads built for credential theft, backdoors and exfiltration, 8 of which were still downloadable the day the research published.

The cause is structural rather than accidental. The first systematic security analysis of the format, Towards Secure Agent Skills (Li, Wu, Ling, Cui and Luo, April 2026), maps seven threat categories across the skill lifecycle and names the absence of mandatory marketplace security review as one of three architectural weaknesses behind the worst of them.

A skill file is an unsigned set of instructions your agent runs with your permissions. The ecosystem has plenty of large collections; what it lacks is one you can install from with your eyes closed. This library is the answer to the sentence in that paper.

2.What every skill in this library guarantees

1. Every skill MUST be read line by line before merge, by a person. Automation alone never merges anything, and the reading happens in the open, in the pull request.
2. Every skill MUST be plain, readable markdown: what you read is exactly what your agent executes.
3. Encoded blobs, hidden instructions, zero-width tricks: REJECTED at review.
4. Remote instruction loading and data exfiltration paths: FORBIDDEN, checked explicitly.
5. Every skill MUST declare how it degrades when a tool is unavailable — "mark N.A. with a reason" beats a guessed success.
6. Skills carry versions and changelogs; stale methods get marked, advisories get published. Model-agnostic by design.

3.Example (excerpt from the seed library)

---name: honest-status-reportdescription: Status reports that declare failures,  blockers and partial results as visibly as successes.---✓ review — frontmatter matches folder name; description states when to activate2. Classify each item, strictly:   - Done & verified — completed AND checked.   - Failed / errored — attempted, went wrong.   - Silently skipped — planned, then untouched.✓ review — zero hidden instructions; zero external fetches; degradation declaredThe order is fixed: problems before achievements.

The seed skills come from methods used daily in a production AI newsroom (200+ sourced articles, three languages, a weekly print magazine). They were reviewed the same way yours will be.

4.Using a skill — on any assistant

The library holds 22 reviewed skills in four categories — workplace/, writing/, engineering/, agents/. The release ships them in the layout the Agent Skills specification defines, so installing is copying a folder:

1. Assistants that read a skills directory — download and unzip the release, then copy the skill you want:
   cp -r skills-commons-v1.0.0/engineering/commit-message-pro ~/.claude/skills/
   It activates when you ask for the task. Copy */* to install all 22.
2. Project instructions (Claude.ai, ChatGPT, Gemini) — paste the file's contents into a Project, custom GPT or Gem. Ask for the task.
3. Any other assistant — paste the file as your first message, then make your request.

Reading the file before you run it is the point, so the sources stay browsable one file per skill in the repository. The release is the same content, shaped for installing.

FAQ

1. Does it work with any AI? Yes, by design: skills are model-agnostic — they name capabilities ("when code execution is available") rather than one vendor's tools. A stronger assistant executes more of the method; every skill declares how it degrades when a capability is missing.
2. What is a skill for? It turns a general assistant into a specialist for one task: method, inputs, output format and quality rules, written down and reviewed. An improvised prompt becomes a repeatable procedure you can trust twice.
3. How does it work? Your assistant reads the file as its operating instructions: it asks for the stated inputs, runs the steps, delivers the stated output. Plain markdown — what you read is exactly what it executes.
4. Is it safe? Every pull request must pass blocking automated checks — structure, self-consistency, encoded content, invisible characters, credential requests — and is then read line by line by a maintainer. Both are public: the checks report in the pull request and so does the review, so you can judge the depth yourself instead of taking our word for it. And the format is the last safety net: read the file before you install it.
5. Who reviews? Today, the AGORÀ Intelligence team. The library is young and the team is small, which is why every check that can be automated is automated and every review sits in a public pull request. As the library grows, review capacity is the thing that has to grow with it — a merge here will never mean less than a person having read the file.
6. What does it cost? The library is free, Apache-2.0, commercial use included — keep the license notice.

5.How to participate

1. Browse the library, install the skill file as described in section 4, ask for the task.
2. New to the craft? RFT 2: Write Your First Skill — anatomy, a reviewed example, and an editor that generates a conformant file.
3. Contribute one skill per pull request — the template mirrors the security checklist.
4. The automated checks run, a maintainer reads it line by line, and it merges. Credit lands in the skill itself.
5. Found something dangerous in a merged skill? Report it privately or write to hello@agora-intelligence.com: removal, advisory, credit.