# AI Act Compliance Skill
> An agent skill for compliance work on **Regulation (EU) 2024/1689 โ the EU Artificial Intelligence Act**. Compatible with **Anthropic Claude Code**, **OpenAI Codex**, and any **AGENTS.md**-aware coding agent (Cursor, Windsurf, GitHub Copilot, Devin, Amp).
>
> **Status: v0.1.0-alpha โ public, usable, seeking validators.** The content is grounded in the Official Journal text and the 2025 Commission guidelines, but external review is wide open. If you are a compliance officer, EU AI law expert, or engineering lead with hands-on AI Act experience: please open issues, suggest corrections, or email `ai@davidemorelli.it`. Acknowledged contributors get credit in the README and a citation when the skill ships v0.1.0 stable.
>
> **Article 50 transparency is in force since 2 August 2026.** Regulation (EU) 2026/1744 (Digital Omnibus on AI, OJ 24 July 2026, in force 27 July 2026) deferred only the Chapter III high-risk dates โ **2 December 2027** for stand-alone Annex III systems and **2 August 2028** for safety components of Annex I products. It did **not** defer Article 50. The single transitional relief is for the Art. 50(2) marking and detection duty on generative systems placed on the market before 2 August 2026, which have until **2 December 2026**. The same Regulation adds **Article 5(1) points (ba) and (bb)**, prohibiting AI systems generating non-consensual intimate imagery ("nudifier apps") and child sexual abuse material, applicable from **2 December 2026**; the new Article 5(1a) makes safeguards relevant only where the prohibited output is a foreseeable outcome, not where it is the system's intended purpose. Article 4 (AI literacy) and the pre-existing Article 5 prohibitions โ in force since 2 February 2025 โ and Chapter V GPAI obligations โ in force since 2 August 2025 โ are unaffected.
## Looking for validators
Specifically, we are interested in feedback on:
- **Article 5 borderlines**: workplace emotion recognition, biometric categorisation carve-outs, real-time RBI law-enforcement exceptions
- **Annex III scope edges**: where high-risk classification is ambiguous in real systems
- **Article 27 FRIA**: who has actually drafted one for a public-sector deployer; what worked, what did not
- **GPAI thresholds**: how the 10ยฒโต FLOPs criterion is being interpreted for fine-tuned and downstream-adapted models
- **Article 50 transparency**: implementation patterns for chatbot disclosure, synthetic-content marking, deep-fake notices
Open an issue at <https://github.com/morellid/ai-act-skill/issues>, or send a private note to `ai@davidemorelli.it`.
## What is this?
A self-contained agent skill that helps compliance teams, AI engineers, and consultants run common tasks on the EU AI Act:
- Classify an AI system or model (in scope? prohibited? high-risk? limited? GPAI? systemic risk?)
- Identify roles (provider, deployer, importer, distributor)
- Check provider obligations for high-risk systems (Articles 8โ22)
- Check deployer obligations including the Fundamental Rights Impact Assessment / FRIA (Articles 26โ27)
- Check GPAI provider obligations (Articles 51โ55)
- Verify transparency obligations under Article 50 (chatbots, synthetic content, deepfakes)
## Installation
The repository **is** the skill. Drop it into your agent's skills directory. The skill follows the `SKILL.md` + frontmatter convention used by both Claude Code and Codex.
| Agent | Skills directory |
|---|---|
| Anthropic Claude Code | `~/.claude/skills/ai-act-compliance/` |
| OpenAI Codex | `~/.agents/skills/ai-act-compliance/` |
### Option A โ install via the helper script (Claude Code / Codex only)
```bash
git clone https://github.com/morellid/ai-act-skill.git
cd ai-act-skill
./install.sh # default: Claude Code
./install.sh --target codex # Codex only
./install.sh --target both # both agents
./install.sh --force --target both # replace existing installs without prompting
```
The script symlinks the repository into the chosen native target(s) by default (so `git pull` upgrades the installed skill in place). Set `INSTALL_MODE=copy` for a snapshot copy.
This helper is for **native Claude Code / Codex skill installs**. If you want to vendor the repo into a project for Cursor, Windsurf, GitHub Copilot, Devin, Amp, or another AGENTS.md-aware harness, use **Option D**. A custom-path run such as `./install.sh /path/to/project/.vendor/ai-act-skill` only copies/symlinks the repo; it does **not** create the project-local `AGENTS.md` pointer or per-agent adapter files.
### Option B โ clone directly into the skills directory
```bash
# Claude Code
git clone https://github.com/morellid/ai-act-skill.git ~/.claude/skills/ai-act-compliance
# Codex
git clone https://github.com/morellid/ai-act-skill.git ~/.agents/skills/ai-act-compliance
```
### Option C โ Codex `$skill-installer`
Inside a Codex session:
```
$skill-installer https://github.com/morellid/ai-act-skill
```
### Option D โ use from another agent (Cursor, Windsurf, GitHub Copilot, Devin, Amp, โฆ)
Many agents read `AGENTS.md` natively (60 000+ repositories already do; the format is governed by the Linux Foundation Agentic AI Foundation). For those agents, vendor this skill into your own project and add a short pointer in your project's `AGENTS.md`:
```bash
# In YOUR project repository:
git submodule add https://github.com/morellid/ai-act-skill.git .vendor/ai-act-skill
```
Then add to your project's `AGENTS.md`:
```markdown
## EU AI Act compliance
When the user asks AI Act questions, walk through the tasks at
`.vendor/ai-act-skill/tasks/` (classify first, then check prohibitions,
then route to provider/deployer/GPAI/transparency). Source extracts at
`.vendor/ai-act-skill/references/extracts/`. Cite articles precisely;
always close with the "support tool, not legal advice" disclaimer.
```
For per-agent adapters (Cursor MDC rule, GitHub Copilot instructions), see [`adapters/`](adapters/) in this repository.
### Option E โ download a release zip
```bash
mkdir -p ~/.claude/skills # or ~/.agents/skills for Codex
curl -L https://github.com/morellid/ai-act-skill/archive/refs/tags/v0.1.0.zip \
-o /tmp/ai-act-skill.zip
unzip -d ~/.claude/skills /tmp/ai-act-skill.zip
mv ~/.claude/skills/ai-act-skill-* ~/.claude/skills/ai-act-compliance
```
### Option F โ install via `npx skills` (vercel-labs CLI)
The [`vercel-labs/skills`](https://github.com/vercel-labs/skills) CLI auto-detects the `SKILL.md` at the repo root and installs it into your agent's skills directory.
```bash
npx skills add morellid/ai-act-skill
```
By default the CLI installs into `~/.claude/skills/`. Use `--list` to preview, or pass an explicit target with the CLI's flags (see `npx skills add --help`).
### Option G โ drag-and-drop on Claude.ai (web)
If you use Claude only via the web (claude.ai, no Claude Code), grab the prebuilt skill zip from the [GitHub Releases](https://github.com/morellid/ai-act-skill/releases) and upload it.
1. Open the latest release: <https://github.com/morellid/ai-act-skill/releases/latest>
2. Download `ai-act-compliance.zip`.
3. Go to <https://claude.ai/customize/skills> and upload (or drag-and-drop) the file.
The zip contains the skill in a single top-level directory (`ai-act-compliance/` with `SKILL.md`, `tasks/`, `references/`, `examples/`, `LICENSE`). Do not extract it before upload.
After installation, restart your agent (Claude Code or Codex) so the new skill is discovered.
### Uninstall
```bash
./uninstall.sh # default: Claude Code
./uninstall.sh --target codex # Codex only
./uninstall.sh --target both # both agents
# or manually:
rm -rf ~/.claude/skills/ai-act-compliance
rm -rf ~/.agents/skills/ai-act-compliance
```
## How to use it (once installed)
In any Claude Code or Codex session, ask questions or give tasks that touch the AI Act:
> "Classify this system under the AI Act: it's a CV-screening tool for HR..."
>
> "Run a FRIA on this high-risk deployment of an emotion-recognition system."
>
> "Check our GPAI model obligations โ we're at 5ร10ยฒโต FLOPs of training compute."
The agent will load this skill, route to the right sub-task, and produce a structured analysis with citations.
In Codex you can also invoke it explicitly:
```
/skills ai-act-compliance
```
## Repository layout
```
ai-act-skill/
โโโ SKILL.md โ the skill entry point (agent reads this)
โโโ README.md โ this file (humans read this)
โโโ LICENSE โ MIT
โโโ CHANGELOG.md
โโโ install.sh โ installer (Claude Code / Codex / both)
โโโ uninstall.sh
โโโ .gitignore
โโโ AGENTS.md โ cross-agent guidance (Codex, Cursor, Windsurf, Copilot, โฆ)
โโโ agents/ โ agent-specific UI metadata
โ โโโ openai.yaml โ Codex display name + default prompt
โโโ adapters/ โ drop-in adapters for non-skill formats
โ โโโ copilot-instructions.md โ GitHub Copilot
โ โโโ cursor-rule.mdc โ Cursor IDE rule
โ โโโ README.md โ how to vendor + adapt
โโโ tasks/ โ detailed sub-task instructions
โ โโโ classify-system.md
โ โโโ check-prohibited-practices.md
โ โโโ check-high-risk-provider.md
โ โโโ check-deployer-obligations.md
โ โโโ check-gpai-provider.md
โ โโโ check-transparency.md
โโโ references/ โ official-source metadata + textual extracts
โ โโโ sources.yaml
โ โโโ extracts/
โ โโโ ai-act-art-3-definitions.md
โ โโโ ai-act-art-5-prohibited.md
โ โโโ ai-act-art-6-9-high-risk-classification.md
โ โโโ ai-act-annex-iii-high-risk-list.md
โ โโโ ai-act-art-26-27-deployer-fria.md
โ โโโ ai-act-art-50-transparency.md
โ โโโ ai-act-art-51-55-gpai.md
โโโ examples/ โ test fixtures
โ โโโ classify-customer-service-chatbot/
โ โโโ classify-hr-emotion-recognition/
โโโ scripts/
โโโ validate.sh โ skill self-check
โโโ fetch-references.sh โ downloads source PDFs and verifies hashes
โโโ build_releases.sh โ builds dist/ai-act-compliance.zip for claude.ai drag-and-drop
```
## Versioning
Semantic versioning, tagged in git (`git tag v0.1.0`). GitHub auto-generates downloadable zips for each tag at `https://github.com/.../archive/refs/tags/<tag>.zip`.
## Validation and tests
The skill ships with two example fixtures (HR emotion recognition, customer service chatbot) under `examples/`, each with a synthetic input and the expected output. To run a structural self-check:
```bash
./scripts/validate.sh
```
For external validation resources (the AI Act is not amenable to fully automated test suites โ compliance is fact-specific), see [`references/sources.yaml`](references/sources.yaml) for:
- The Commission's official Compliance Checker at the AI Act Service Desk
- The artificialintelligenceact.eu open-source tracker
- Harmonised standards: EN 18286:2026 (QMS) referenced by Commission Implementing Decision (EU) 2026/2274 (OJ 14 October 2026); further standards under preparation by CEN-CENELEC JTC 21
## Contributing
Issues and PRs welcome. Two non-negotiable principles:
1. **Only official sources** โ every normative claim must trace back to the Regulation, an EU Commission delegated/implementing act, an EU AI Office guideline, an EDPB joint guideline, or a published harmonised standard. Hashes in `references/sources.yaml`.
2. **Disclaimer always present** โ the skill is a support tool, never a substitute for legal counsel or notified body assessment.
## Acknowledgements
- **Richard B.** ([LinkedIn](https://www.linkedin.com/in/richard-b-42469439b/)) โ flagged the Digital Omnibus on AI timeline change (Commission proposal 19 November 2025; Council and Parliament alignment in March 2026) before v0.1.0 stable, prompting the pending-delay disclosure pass across the skill and the landing page.
## Licence
MIT โ see [LICENSE](LICENSE).
## Disclaimer
This skill is provided for informational and compliance-support purposes. It does not constitute legal advice, does not replace assessment by a qualified lawyer, and does not replace conformity assessment by a notified body where required. Penalties under Article 99 of the AI Act reach EUR 35 million or 7% of total worldwide annual turnover for the most serious violations. Use at your own risk.
README-Quelldatei lesen