Guide · Agent skill
How to use the AI Impact Assessment Generator agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The aia-generation skill is an agent skill for Claude Code or a compatible agent that helps a legal or governance team document a new AI system. It asks the user a conversational set of questions about the system, classifies it against the AI regulations in the practice's configured regulatory footprint, checks it against the company's AI policy, and writes the result as an assessment in the team's house format. After this guide you will know what the skill does, how to install it, how to trigger it, and what it needs before it can produce a useful result.
AI Impact Assessment Generatoranthropics/claude-for-legal · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Download the aia-generation skill ZIP from this page. It contains SKILL.md.
- 2Unzip into a skills folder. Unzip into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder for one project.
- 3Populate the practice CLAUDE.md. Fill in the impact assessment house style and regulatory footprint in ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md.
- 4Connect a legal research tool. Make sure the agent can reach a legal research tool such as Westlaw or EUR-Lex for regulatory lookups.
- 5Describe the use case. Ask in plain terms, for example 'Run an AIA for our résumé screening tool', or run the slash command with a short description.
- 6Answer the intake questions. Respond to the conversational intake about what the system does, who it affects, and what oversight exists.
- 7Review the output and verify citations. Check the assessment, the conditions list and the handoff flags, and verify every tagged citation before a lawyer signs off.
What does the skill make the agent do?
An AI impact assessment is a documented decision about whether an AI system is designed and deployed responsibly. The skill walks the agent through that decision in order: a structured intake about what the system does, how it reaches its outputs, who it affects if it is wrong, and what oversight exists; a risk analysis; a regulatory classification for each regime in scope; a policy consistency check; and a recommendation with conditions.
The output is a written assessment in the team's house format, plus a list of conditions and handoff flags for related reviews such as a privacy impact assessment or a vendor review. The skill is a legal-team workflow tool. The lawyer still makes the final decision.
- Intake: a conversation, not a form
- Regulatory classification per regime in the configured footprint, with citations
- Policy diff against the company's AI policy commitments
- Output: assessment document, conditions list, and handoff flags
Install the aia-generation skill
Download the skill ZIP from this page. The skill's only file is SKILL.md. Unzip it into one of two locations: your personal skills folder at ~/.claude/skills/, or a project's .claude/skills/ folder if you want the skill available only within that project.
If you use the Claude apps instead of Claude Code, upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
~/.claude/skills/Set up the practice configuration before you run it
The skill depends on a practice-level CLAUDE.md file at ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/. That file should contain your impact assessment house style and a populated regulatory footprint. The skill reads these sections: ## Impact assessment house style, ## Regulatory footprint, ## Use case registry, ## Governance tiers, ## Who's using this, and the AI policy commitments. Without them, the skill will not produce a jurisdiction-accurate result.
The skill also needs access to a legal research tool, such as Westlaw or EUR-Lex, for regulatory lookups. Web search is optional and is used only with your approval.
- Required: Claude Code or a compatible agent with the claude-for-legal ai-governance-legal plugin configuration
- Required: practice-level CLAUDE.md with house style and regulatory footprint populated
- Required: access to a legal research tool for regulatory lookups
- Optional: web search, used only with user approval
~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.mdHow do you trigger the skill with natural requests?
You can invoke the skill with the slash command or simply describe what you need. The skill's description lists these phrases as triggers: 'impact assessment for', 'assess this AI use case', 'run an AIA', 'generate an AIA', 'we need to document this AI system', and 'AI risk assessment for X'. It also runs when you follow a conditional triage result.
Give the use case in plain terms. The more concrete the description, the fewer intake questions the agent needs to ask.
- 'Run an AIA for our résumé screening tool.'
- 'We need to document this AI system: a chatbot that answers customer billing questions.'
- 'Assess this AI use case: automated scoring of loan applications.'
/ai-governance-legal:aia-generation "AI résumé screening for HR"What happens during a run?
The agent first reads the practice CLAUDE.md and confirms the house style is populated. It then chooses a risk track, either fast or full, based on the governance tier and the use case characteristics. Next comes the intake, followed by regulatory classification for each regime in the footprint. The agent then writes the assessment in house style, runs the policy diff, and produces the outputs.
If a research query returns few or no results for a regime, the skill stops and reports what it found. It does not fill the gap from web search or model knowledge without asking you first. You choose whether to broaden the query, try another tool, use web search, or mark the item as unverified.
- Step 1: confirm house style and choose the risk track
- Step 2: intake through conversation
- Step 3: regulatory classification with primary-source citations
- Step 4: write the assessment in house style
- Step 5: policy diff against AI policy commitments
- Step 6: output assessment, conditions and handoff flags
How do citations get tagged for verification?
Every citation in the assessment carries a source tag so you can see how much checking it needs. [settled] marks stable, well-known references. [verify] marks real model-knowledge citations that should be checked, such as specific regulator guidance or effective dates. [verify-pinpoint] marks pinpoint citations like article numbers, which carry the highest fabrication risk and must be checked against the primary source. Tool-retrieved citations keep their tool tag, such as [Westlaw] or [EUR-Lex]. Web-search citations stay [web search — verify], and citations you supply stay [user provided].
The tags are never removed. Their purpose is to show you where the real verification work is.
- [settled]: lower verification priority
- [verify]: check before relying on it
- [verify-pinpoint]: always check against the primary source
- [Westlaw] / [EUR-Lex]: retrieved by tool, keep the source tag
- [web search — verify] and [user provided]: keep as tagged
Safety notes and tips
The skill runs no scripts. It does use the network, because it queries legal research tools and may use web search with your approval. The assessment is only as current as the sources it can retrieve, so treat it as a first draft.
If you work with matters rather than practice-level context, the skill checks the matter-workspace settings. Outputs for a matter are written to the matter folder under the plugin's matters/ directory. It does not read other matters' files unless cross-matter context is turned on.
- Scripts: none. Network: yes, for research and approved web search
- Verify every [verify-pinpoint] citation against the official text
- Confirm the deployment is inside the configured jurisdictions; the skill warns that its analysis may not apply outside them
- Check the matter switch before running: /ai-governance-legal:matter-workspace switch <slug>
/ai-governance-legal:matter-workspace switch <slug>What people use it for
- Documenting a new hiring screening tool, such as AI résumé screening for HR, in the team's house format
- Recording a customer-facing AI decision process so the file shows its classification and conditions
- Producing a consistent, citation-tagged assessment for compliance staff reviewing several AI use cases
- Creating a first draft that a lawyer then verifies against primary sources
- Generating handoff flags so a privacy impact assessment or vendor review starts in parallel
Frequently asked questions
Is the aia-generation skill free to use?
The source does not state pricing or licence terms for using the skill beyond its licence, which is Apache-2.0. You still need Claude Code or a compatible agent and access to a legal research tool such as Westlaw or EUR-Lex, and those have their own terms.
What do I need before I can install and run it?
You need Claude Code or a compatible agent with the claude-for-legal ai-governance-legal plugin configuration, a practice-level CLAUDE.md at ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/ with house style and regulatory footprint populated, and access to a legal research tool. Without the configuration file, the skill will not produce a jurisdiction-accurate result.
Where do I put the skill files?
Unzip the downloaded ZIP into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
Can the skill give me final legal sign-off?
No. The skill's citations are partly from model knowledge and are tagged for verification, so a qualified lawyer must check them. The skill is designed to produce a first draft, and the lawyer makes the final decision.
What if the skill's research comes back with few results?
The skill stops and reports the number of results and the coverage gap. It offers options: broaden the query, try a different research tool, use web search with results tagged [web search — verify], or flag the item as unverified. It does not fill the gap silently.
How is this different from a privacy impact assessment?
A privacy impact assessment asks whether personal data is handled lawfully. An AI impact assessment asks whether the AI system is designed and deployed responsibly. The two often run in parallel and are not substitutes for each other, so the skill flags a PIA as a handoff where needed.
Does the skill run scripts or use the internet?
It runs no scripts. It does use the network, for legal research lookups and, only with your approval, web search.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are an in-house legal or governance team that needs a consistent first draft of an AI impact assessment, and you have a populated practice CLAUDE.md and a legal research tool. Pick something else if you need final legal sign-off from the tool alone, if your organisation has no configured regulatory footprint, or if your deployment falls outside the jurisdictions you have configured.
Download the skill