Many of these came from GRC people under the launch post. Ask yours on GitHub if it is not here.
An open library of 60 GRC skills for AI agents. Each skill writes down how an experienced person does one piece of governance, risk and compliance work: a scope statement, a first risk register, an internal audit checklist, a vendor triage, a DPIA screening.
Your agent follows the skill on your own files and drafts the work. You make the calls.
Consultants, auditors, GRC leads and anyone who keeps a compliance programme running. It is written for people who already know the work and want the gathering, drafting and cross-checking off their plate.
No. A GRC expert is paid for the calls they make, and those stay with a person: the scope, the risk you accept, what is effective, what you sign. Every skill has hard rules that stop the agent before any of those calls and hand them back to you.
The skills are working aids. They are not legal advice and not an audit opinion.
Nothing. The library is MIT licensed and there is nothing to sign up for. You run it with the agent you already pay for.
The optional framework lookup through ISMS Copilot works on every ISMS Copilot plan, Free included, and its questions count against that plan. Free runs fast mode only, 10 messages per 4 hours; the think mode the skills ask for needs a paid plan.
Skills are tagged with the frameworks they relate to: ISO 27001, ISO 27701, ISO 42001, SOC 2, PCI DSS, GDPR, NIS 2, DORA, the EU AI Act, the Cyber Resilience Act, HIPAA, NIST CSF, plus general skills that fit any programme.
A skill is one piece of work, not one framework. A risk register skill serves ISO 27001 and SOC 2 alike.
Each skill names its author, and every author has a page on this site. The first skills are by Tristan Roth. Before a skill goes in, a GRC specialist model challenges it: would a practising auditor follow this, does anything misstate a framework, what is missing.
No. Framework names appear only as tags saying which framework a skill relates to, and they belong to their owners. grcskills is not affiliated with, endorsed by or sponsored by any standards body or regulator.
Claude Code, Codex, Cursor and any agent that reads SKILL.md. Every skill is a plain SKILL.md file in the Agent Skills format.
Open your agent in the folder that holds your compliance material and paste this: "Read grcskills.com/llms.txt, then run the grc-setup-review skill on my GRC setup: rank the 10 skills that would help me most, list what is broken or missing, and write a 30-day plan."
It reads your policies, registers and exports without changing them, then writes four drafts in a grc-setup-review folder: an inventory, the top 10 skills for your setup, what looks broken, duplicated or missing, and a 30-day plan. On the fictional company in the README demo, the run took about 4 minutes.
With the skills CLI: npx skills add better-isms/grcskills#v2026.10.20.3 for the whole library, or add --skill and a skill name (for example --skill risk-register-first-pass) for one skill. In Claude Code you can also run /plugin marketplace add better-isms/grcskills, then /plugin install grcskills@grcskills. Or clone the release: git clone --branch v2026.10.20.3 https://github.com/better-isms/grcskills.
Every install path is pinned to a release tag, never to the main branch.
Yes. Download the skill folder from the release on GitHub and put it in your agent's skills folder. To publish a skill without git, your agent can file a "Submit a skill" issue instead of a pull request.
Yes. The library is MIT licensed: copy it, adapt it and use it in paid client work. Keep the licence notice. If your version is better, consider publishing it under your name.
The skills themselves send nothing anywhere, apart from the optional ISMS Copilot question, which you see first. They work on your files with your agent and write drafts in your folder.
What your agent reads still goes to the AI model it runs on, so the risk is the one you already accepted for that agent. If you want to test first, run it on a copy with no client data.
The setup review only reads your material. Each skill writes its drafts only inside its own folder (for example ./risk-register-first-pass/) and asks before overwriting anything there. No skill may write to shared systems such as compliance platforms, tickets or repositories without an explicit approval step from you.
No skill marks a control effective, accepts a risk or signs anything on your behalf.
No. A skill may never upload or send data, download or run code, read credentials or link outside an allowlist (official law, regulator and standards sites, plus this project on GitHub). The validator rejects what it can check mechanically on every pull request, and maintainers read every file before a release.
The one prompt makes two network calls, both visible to you: it reads grcskills.com/llms.txt and clones the pinned release into a .grcskills folder. The optional ISMS Copilot lookup is the only other one, and you see each question first.
Installs come from release tags. Each release is immutable and carries a build provenance attestation. An outside pull request may only change its author's own folder, and the founder approves every outside merge.
If a skill does something it should not, report it privately through a GitHub security advisory. A confirmed bad skill is reverted and a new release is cut.
No. Skills never restate what a standard requires and never quote official control titles. When a framework question comes up, the skill tells the agent to check the official text, or to ask ISMS Copilot if you connected it. Get the official texts from their publishers.
There is a skill for exactly this. The evidence freshness check reads the date inside each artefact (the clock on the screenshot, the report header), never the file date, and flags what is stale, undated or outside the audit period. Run it four to eight weeks before fieldwork, early enough to fix gaps by running the activity.
Not automatically today. You rerun the document consistency check: it reads policies, procedures and registers side by side and lists every place two of them disagree, with a fix for each owner.
A lot of it is. That is why the drafts are not a summary to sign. Many skills cite the file or answer behind each finding, and the gap assessment and risk skills mark what is claimed but not evidenced, so the expert goes into the evidence where it matters.
Each draft is a plain file in your folder, and many skills cite the file behind each finding. Keep those drafts in the same version control as your evidence, and add your decision next to each one. grcskills does not run a tracking system of its own today.
Some of them, like access review prep or restore test evidence, are first line work. They are in because GRC people often end up chasing that evidence before an audit. The skill prepares it. The owners in IT or IAM still do the review.
Not today. Skills never download or run code, so none ships a scanner. Some skills also stay out of the data on purpose: the health data flow mapping skill never opens patient records, even to confirm a field holds health data. It maps where health data enters, moves and rests from schemas, field names and configuration, and calls out the copies nobody designed (logs, support tickets, test databases) for an owner to check.
Same team. grcskills is maintained by Better ISMS, the company behind ISMS Copilot. The skills work fully without it.
If you connect ISMS Copilot to your agent, a skill can ask it a framework question. The skill sends only a short question in its own words, shows it to you first, and never includes file contents, secrets, personal data or evidence.
Add one URL to your AI: https://account.ismscopilot.com/v1/account/mcp. In claude.ai, Claude Desktop, Claude Code and ChatGPT you sign in to ISMS Copilot and click Approve. Cursor, Codex and other clients use a personal access token you create in the ISMS Copilot app. Nothing to install.
It works on every ISMS Copilot plan, Free included; Free runs fast mode only, and the think mode the skills ask for needs a paid plan. You can disconnect it any time in Settings, Connected apps.
Releases are dated tags such as v2026.10.20.3, and every change is in the changelog. Your install stays on the tag you chose until you install a newer one. The one prompt always reads the current release from llms.txt.
Tell us. Open an issue on GitHub with what the skill did and what it should have done. For a security problem, use a private security advisory instead.
Yes, and it carries your name. Paste the contribution prompt into your agent: it interviews you about one GRC job you do well, drafts the skill in the library format, validates it, shows you every file and opens the pull request once you say yes.
Your author page shows your name, title, LinkedIn profile and company domain, never an email address.
Yes, at any time. Open an issue, or a private security advisory if it is sensitive. A removed author's profile and skills leave the next release.
Automatic checks first, then a security review, a check that no standards text slipped in, a framework accuracy challenge and a review by a second model that did not write the skill. The founder approves every outside merge.
Open an issue on GitHub. If the answer helps others, it goes on this page.