Skip to content

oliverschmidtprietz/EU-AI-Act-High-Risk-Classifier

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EU AI Act High-Risk Classifier — Deployment Guide

See CHANGELOG.md for version history.

Overview

ai-act-high-risk is a depth assessment skill for EU AI Act Art. 6 high-risk classification (Regulation (EU) 2024/1689), grounded in the European Commission's draft Art. 6(5) classification guidelines (general principles + Annex I + Annex III) published for stakeholder consultation in 2026.

Given an AI system description and intended-purpose evidence, the skill applies a 5-step decision tree and produces:

  • Artefact 1 — structured decision block (terminal verdict with citations)
  • Artefact 2 — practitioner memo (1–2 page narrative addressed to a DPO / AI officer)
  • Artefact 3 — JSON interchange artefact (machine-readable hand-off for downstream skills)

What this skill does

  • Art. 6(1) + Annex I — assesses whether the AI system is itself a regulated product or a safety component of one under Union harmonisation legislation, applies the Art. 3(14) two-prong test (intent-based safety function and consequences-based failure prong), and determines whether a third-party conformity assessment regime is triggered.
  • Art. 6(2) + Annex III — screens against the eight Annex III high-risk areas (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice/democracy), applies the Art. 6(3) exception filter and profiling re-exception.
  • Art. 25 substantial-modification trap — flags when fine-tuning, rebranding, or repurposing a non-high-risk system (including GPAI) flips an enterprise into provider status with full Chapter III obligations.
  • AI Omnibus 2026 timeline — applies the postponed effective dates throughout (Annex III: 2 December 2027; Annex I: 2 August 2028; Art. 111(2) legacy cut-off: 2 December 2027).

File structure

EU-AI-Act-High-Risk-Classifier/
├── SKILL.md                                       # Main skill instructions (deploy this)
├── CHANGELOG.md                                   # Version history
├── LICENSE.txt                                    # AGPL-3.0
├── evals/
│   └── evals.json                                 # 10 test cases, 60 assertions
└── references/                                    # 16 grounded reference files
    ├── art-6-general-principles.md                # Commission's general principles (¶1–14, ¶448–451)
    ├── art-6-1-annex-i-guidelines.md              # Annex I section (¶15–62)
    ├── art-6-2-annex-iii-guidelines.md            # Annex III section (¶63–447)
    ├── safety-function-checklist.md               # Art. 3(14) safety-function taxonomy (¶37 box)
    ├── annex-i-section-a-vs-b.md                  # Section A (NLF-based) vs Section B mapping
    ├── annex-iii-area-1-biometrics.md             # Annex III Area 1 — Biometrics
    ├── annex-iii-area-2-critical-infrastructure.md
    ├── annex-iii-area-3-education.md
    ├── annex-iii-area-4-employment.md
    ├── annex-iii-area-5-essential-services.md
    ├── annex-iii-area-6-law-enforcement.md
    ├── annex-iii-area-7-migration.md
    ├── annex-iii-area-8-justice-democracy.md
    ├── art-6-3-exception-decision-tree.md         # Art. 6(3) exception filter with worked examples
    ├── art-25-substantial-modification-flag.md    # Art. 25 quasi-provider trap
    └── ai-omnibus-timeline-postponements.md       # AI Omnibus 2026 postponement citation chain

Deployment

Claude.ai (User Skills)

  1. Open Settings → Profile → Custom Skills (or equivalent Claude.ai skills entry point).
  2. Upload the entire repo contents (preserve SKILL.md at the root and the references/ and evals/ subdirectories).
  3. The skill will auto-trigger on phrases such as "assess high-risk under the AI Act", "check Annex I", "check Annex III", "is our AI system high-risk", "Hochrisiko-Einstufung", "Sicherheitsbauteil", "Anhang III Nr. X".

Claude Code (~/.claude/skills/)

git clone https://github.com/oliverschmidtprietz/EU-AI-Act-High-Risk-Classifier.git
mkdir -p ~/.claude/skills
ln -s "$(pwd)/EU-AI-Act-High-Risk-Classifier" ~/.claude/skills/ai-act-high-risk

(Or cp -r the folder if you prefer not to symlink.) Claude Code will discover the skill on next start.

Claude API / Anthropic SDK

Load SKILL.md and the relevant references/ files into your context as needed. The skill's structure assumes progressive disclosure: SKILL.md is the entry point; reference files are loaded when the decision tree points to them.

Usage

Trigger phrases

  • "Assess high-risk status under the AI Act"
  • "Check Annex I / Annex III"
  • "Apply the Art. 6 classification"
  • "Check Art. 6(3) exception"
  • "Is our AI system high-risk?"
  • "Perform safety component analysis"
  • "Check third-party conformity assessment trigger"
  • DE: "Hochrisiko-Einstufung", "Hochrisiko-KI-System", "Sicherheitsbauteil", "Anhang I", "Anhang III", "Anhang III Nr. X"

Sample prompt

"We are a lift manufacturer. We have integrated an AI system that manages door-closing timing and obstacle detection in our passenger lifts. The provider's documented intended purpose is efficient lift operation and user experience. The lift is covered by Directive 2014/33/EU on lifts (Annex I Section A). The lift requires third-party conformity assessment under that directive. Run the high-risk classification."

The skill will walk the 5-step decision tree, apply the Art. 3(14) failure-or-malfunctioning prong, anchor to Commission Guidelines Annex I ¶46 (which uses lifts as a worked example), produce the three artefacts, and route on to ai-act-roles / ai-act-obligations for downstream workflow.

Capabilities summary

Feature Description
5-step decision tree AI gating → intended-purpose framing → Art. 6(1)+Annex I → Art. 6(2)+Annex III → Art. 25 trap
Safety-component two-prong test Art. 3(14) intent-based AND consequences-based prongs, both alternatives sufficient
Annex I Section A/B distinction NLF-based legislation (full Chapter III) vs Section B legislation (Art. 6(1) + 102–109 + 112 only)
8 Annex III area files Biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice/democracy
Art. 6(3) exception filter Four limbs (a)–(d) with worked examples and profiling re-exception
Art. 25 substantial-modification flag (a) rebranding, (b) substantial modification, (c) repurposing non-HR system (incl. GPAI) into HR use
AI Omnibus 2026 timeline Postponed dates applied throughout (Annex III: 2 Dec 2027; Annex I: 2 Aug 2028; Art. 111(2): 2 Dec 2027)
Commission-guideline paragraph anchors ¶27, ¶37, ¶46, ¶48-49, ¶170 etc. cited explicitly in outputs
JSON interchange artefact Machine-readable verdict block for downstream cross-skill workflow

Regulatory basis

Document Reference
EU AI Act Regulation (EU) 2024/1689 (esp. Art. 3, 6, 25, 49, Annex I, Annex III)
Commission draft Art. 6(5) classification guidelines General principles, Annex I, Annex III (issued 2026 for stakeholder consultation)
AI Omnibus 2026 Regulation amending Art. 113 effective dates and Art. 111(2) cut-off
Annex I underlying acts Machinery Reg. 2023/1230, Lifts Dir. 2014/33/EU, Toys Safety Reg., RED 2014/53/EU, MDR/IVDR, etc.
Decision 768/2008/EC Conformity assessment modules A–H1 referenced in Art. 6(1) third-party trigger analysis

Benchmark (v1.0 release)

/skill-creator iteration-1 sweep — 10 cases × 60 assertions, 1 run per configuration:

Metric With skill No-skill baseline Δ
Pass rate 91.3% ± 16.4% 70.7% ± 27.9% +20.6pp
Wall-clock 107.2s ± 30.0s 123.7s ± 29.5s −16.5s
Tokens 43,372 ± 4,714 29,311 ± 1,898 +14,060

The skill's value-add concentrates in (a) post-cutoff AI Omnibus dates (baseline missed 4 of 4 date assertions; skill: 4 of 4 correct) and (b) Commission Guidelines paragraph anchors that baselines without the 2026 guidelines cannot reproduce.

Disclaimer

This skill provides structured high-risk classification guidance based on the EU AI Act (Regulation (EU) 2024/1689) and the Commission's draft Art. 6(5) classification guidelines issued for stakeholder consultation. The Commission guidelines are non-binding; authoritative interpretation rests with the Court of Justice of the EU. This is not legal advice. Final classification decisions should involve qualified legal counsel with AI Act expertise.

License

AGPL-3.0 — see LICENSE.txt.

Related skills

  • ai-act-classifier — broad risk-tier triage; routes to this skill for high-risk depth.
  • ai-act-knowledge — Q&A over the full AI Act + Commission guidelines.
  • ai-act-roles — provider / deployer / importer / distributor determination, including Art. 25.
  • ai-act-obligations — once high-risk verdict is in, maps obligations by role + risk tier.
  • ai-act-report — formal compliance report generation.
  • ai-act-quick — 15–25 min preliminary assessment; routes here for full Art. 6 depth.

Created by Oliver Schmidt-Prietz — canonical source: oliverschmidtprietz/claude-skills (private monorepo).

About

EU AI Act high-risk classification — Art. 6 depth assessment skill for Claude, grounded in Commission draft Art. 6(5) guidelines (general principles, Annex I, Annex III). AGPL-3.0. Canonical source: oliverschmidtprietz/claude-skills.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors