Ansvar AI docs
    ansvar.eu →

    Agent skills

    A skill is a single markdown file you give your AI agent once. It carries a complete compliance workflow — which gateway tools to call, in what order, with which arguments, and the grounding rules that make the output citable — so the agent applies it in every conversation without you pasting instructions. Skills are the packaged form of Instruct your agent: the same cite-or-refuse discipline, plus a task-specific method.

    The files are free to download and licensed CC BY 4.0. Running the compliance skills needs the gateway connector and an Ansvar account; the engineering practice skills further down need neither. Start with Using Ansvar, the routing skill: it carries the grounding method alone — route, scope, fetch, cite, refuse — and exists because a model without it answers compliance questions from training data whenever nobody says “Using Ansvar” in the message. Installed once, it makes routing the standing instruction instead of a per-message reminder. The task skills below assume its method and each add a deep, task-specific workflow. The Using Ansvar, CRA, and Incident Reporting skills use only tools available on every plan, and the calls each one instructs were verified against production on a free-tier connection before publication. The Regulatory Threat Model skill runs the STRIDE workflow, which every plan can start within its monthly allowance — one run on Free, two on Solo, five on Premium; the LINDDUN privacy analysis it offers second needs Premium. Below the task skills, one skill per workflow family carries the method the gateway's workflow engine itself runs.

    Available skills

    Using Ansvar — the routing skill

    “Using Ansvar: do we need a DPIA for this feature, and which Swedish rules add to the GDPR here?”

    • The install-once form of the routing habit: standing instructions that route bare compliance questions through the connector instead of letting the model answer from training data — no per-message reminder. Built for the models that otherwise skip the connector.
    • Scopes every search to a jurisdiction or framework, reduces the question to legal key terms in the language of the law being searched, and, when the gateway reports withheld relaxed matches, offers the user a broadened re-run before treating anything as unanswered.
    • Quotes from fetched provision text with a source URL on every claim, and re-validates citations reused from earlier conversations.
    • Refuses rather than invents: when the served sources do not answer, it names the searches it ran and says so.

    Get it: SKILL.md (the file itself) · ready-made ZIP (for the claude.ai skill upload).

    CRA Vulnerability & Reporting Obligations

    “Using Ansvar: what does the EU Cyber Resilience Act mean for our product — and what does this CVE trigger legally?”

    • Determines CRA scope and product classification from the served regulation and implementing-act text, never from a category guess.
    • Branches the duties by role — manufacturer, importer, distributor, open-source steward — and fetches the application and transitional dates before calling any duty live.
    • Joins live CVE, CISA-KEV, and EPSS intelligence to the CRA's actively-exploited-vulnerability test, applied limb by limb.
    • Walks the full Article 14 reporting ladder — early warning, notification, final report, user communications — quoting deadlines verbatim with their trigger events.

    Get it: GitHub repository (canonical, with README and license) · SKILL.md (the file itself).

    Incident Reporting Navigator (EU)

    “Using Ansvar: we have an incident. Who must we notify, in which member state, and by when?”

    • Screens one incident across NIS2, GDPR, DORA, and the CRA — per legal entity, with each entity's own roles and timestamps.
    • Checks every duty was actually in force for the incident before reporting it, keyed to the duty's own trigger time.
    • Resolves the receiving authority per regime and member state from served national law — Dutch, German, and other transpositions — or says plainly that the name is unresolved.
    • Produces a deadline table in which every duty, authority, and date carries an official-publisher citation.

    Get it: GitHub repository (canonical, with README and license) · SKILL.md (the file itself).

    Regulatory Threat Model

    “Using Ansvar: we shipped an app fast with an AI coding agent. What can go wrong, and what does the law actually require us to check?”

    • Runs a STRIDE threat model and a LINDDUN privacy threat analysis from a plain-language system description — no document upload required.
    • Checks the system's dependencies against live CVE, CISA KEV, and EPSS data.
    • Maps the findings to the regulatory obligations that apply — GDPR Art. 25/32, NIS2 Art. 21, the Cyber Resilience Act, the AI Act — cited to the served legal text.
    • Marks a requirement unresolved instead of guessing when none of the enrichment passes returns a citation.
    • Runs the STRIDE threat model within every plan's monthly allowance — 1 run on Free, 2 on Solo, 5 on Premium. The LINDDUN privacy analysis needs Premium.

    Get it: GitHub repository (canonical, with README and license) · SKILL.md (the file itself).

    ISO Standards Expert

    “Using Ansvar: what does ISO/IEC 27001 require for a risk treatment plan, and what should ours contain?”

    • Fetches licensed clause text per standard — SS-EN ISO/IEC 27001:2023, 27002:2022, 27005:2024, 42001:2026 and SS-ISO/SAE 21434:2021 — through the ISO Standards add-on, purchasable on any plan including Free.
    • Quotes every excerpt verbatim with the SIS copyright notice and source URL; its own analysis is labelled as analysis, never blended into the standard's text.
    • Serves selected parts only, never the complete standard — and reports an unevaluated-scope list when the licence coverage ceiling withholds a fetch, instead of papering over the gap.
    • Without an add-on it answers ISO 27001-shaped questions from the free Secure Controls Framework cross-reference, labelled as such on every result.

    Get it: GitHub repository (canonical, with README and license) · SKILL.md (the file itself).

    Workflow family skills

    A workflow is a run the gateway drives stage by stage — it asks for what it needs, fetches the law behind every requirement, and hands back a finished report. These skills are the agent's half of that run, one per family: how to open a run and scope it, what each stage expects, the evidence and citation rules the output has to meet, how to relay the delivery receipt the server returns, and starter questions to paste. Install the family you work in — they sit alongside the routing skill above rather than replacing it.

    • Regulatory Gap Analysis — Assess an organisation against a regulatory framework control by control, with article-level citations and a coverage score. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Threat Modelling — STRIDE security threat models and LINDDUN privacy threat models over a described system, with scored threats and mapped controls. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Impact Assessments — DPIA and FRIA — GDPR Article 35 data protection impact assessments and EU AI Act Article 27 fundamental rights impact assessments, with the national Article 35(4) lists and the Charter rights in scope. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Threat Analysis and Risk Assessment — Structured risk assessments with likelihood and consequence bands, scored against declared thresholds and closed with treatments — including automotive, rail, robotics and OT TARA. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Tender Review and Audit — Bidder-side completeness review of a public tender, and buyer-side lawfulness audit of the tender itself, each requirement cited to its legal basis. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Vulnerability Decisions — Contextual scoring of an already-scanned finding set with KEV and EPSS, a costed control-investment plan, deferral records for what stays open, and a merged CycloneDX VEX. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Adversary Tabletop — A crisis simulation run turn by turn against a scripted adversary, scored from the decisions the team actually made, with an after-action report. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Document Review — Review an uploaded document against a regulatory baseline, with every finding anchored to a paragraph-level citation and a content hash. Get it: SKILL.md · ready-made ZIP · version 1.0.4.
    • Drone Operations — SORA — A SORA determination chain for a specific-category UAS operation: ground and air risk, SAIL, the applicable operational safety objectives with robustness levels, and containment. Get it: SKILL.md · ready-made ZIP · version 1.0.4.

    These files are not written for the website. They are generated from the workflow engine's own instruction library and imported here by checksum: the published file, the release manifest that hashes it, and the pinned build are checked against each other in CI, so a skill on this page cannot drift from the workflow it describes. The set above was built against regulatory content bundle 1.5.0.

    Engineering practice skills

    Not every skill needs the gateway. The affect pack is the engineering practice we run on our own agents: at the end of any substantial coding, review, or audit task, the agent files exactly one finding — or an explicit null — through the lens of a specific emotion. Frustration finds what's broken; embarrassment for the user finds what works but shouldn't ship; protectiveness finds the load-bearing weirdness a cleanup pass would delete. No account, no connector, nothing to configure.

    Agent Affect Skills — whine, cringe, protect

    “What did your agent notice that it had no channel to tell you?”

    • agent-whine — one structured complaint per task: friction, hidden coupling, specs that contradict themselves. The subjective tone is the signal, and a complaint that turns out to be wrong still marks where a competent reader got confused.
    • agent-cringe — pictures one specific person using what just changed and reports where it works but you'd want to look away, with the smallest concrete fix.
    • agent-protect — defends code that looks wrong but must not change, naming the concrete failure if someone deletes it naively.
    • agent-affect-checkin — runs all three as one end-of-task pass; the single skill to wire into review and audit agents.

    Get it: GitHub repository (canonical, with README and license) · the files themselves: check-in · whine · cringe · protect.

    Install

    Claude (claude.ai): upload the skill as a ZIP via the Skills section in Settings. Turn on code execution in Settings first — the Skills upload only appears once it is on. Using Ansvar and every workflow family skill have a ready-made ZIP above — download it and upload it as-is; for the remaining skills, zip the skill folder yourself. For the compliance skills, add the gateway under Settings → Connectors if you have not already:

    https://gateway.ansvar.eu/mcp

    Claude Code: place the folder under .claude/skills/ in your project, with the gateway configured as an MCP server.

    ChatGPT, Copilot, Gemini, and other MCP clients: attach SKILL.md as standing instructions for the conversation or project, with the gateway connected — or start smaller: the dedicated setup guides for supported clients each carry the routing pack (the compact instruction block the Using Ansvar skill packages) and say where that client keeps standing instructions. Client-by-client setup is in Setup.

    Why these are safe to rely on

    • Nothing from model memory. Every legal claim the agent makes is fetched from official publisher text at answer time and cited with a source URL. Where the sources do not answer, the skill instructs the agent to say so — a connector failure is never converted into a legal conclusion.
    • Reviewed before publication. Each compliance skill under Available skills went through a two-round adversarial legal-accuracy review with live EUR-Lex cross-checking before release. The workflow family skills carry the same instructions the workflow engine runs, and reach this page only through the checksummed import described above.
    • Re-verified nightly. The exact calls the CRA, Incident Reporting, Regulatory Threat Model, ISO, and Using Ansvar skills instruct are pinned in the gateway's production verification battery and re-checked against the live service every night — if serving ever drifts from what a published skill expects, we find out before you do.