A governance and delivery asset library for AI agents — skills, specs, protocols and rules. See mission and vision.
When an agent works inside this repository, AGENTS.md is authoritative for the execution contract, the four asset layers, precedence, discovery and skill matching. Terminology is defined in docs/architecture/terminology.md.
68 skills (index), 31 rules (index), 20 specs (index) and 1 protocol (index).
| Area | Count | Representative skills |
|---|---|---|
| Governance & planning | 17 | Deriving mission → vision → North Star → strategic goals → roadmap layer by layer; backlog scoring, dependency mapping, promotion and archival; plan-next diagnoses what to do next |
| Code review | 27 | Pre-coding artifact reviews, post-coding implementation alignment, and orchestrate-code-review across 8 languages, frameworks, libraries and seven engineering concerns |
| Delivery & release | 9 | Commits, worktree delivery and integration, release package build and publication, announcements, test execution, local redeployment |
| Docs & assets | 5 | Generating README, AGENTS.md and GitHub Actions; refining skill design; decontextualizing text |
| Integration & ops | 4 | NATS cross-team messaging, macOS Keychain credential management, agent test scaffolding |
| Research & opportunity | 6 | Open, policy, market and competitive research; internal opportunity assessment; product opportunity package |
Skills are callable from Claude Code, Cursor, Codex and 20+ other agents. For a worked end-to-end flow, see the roadmap planning guide; to find an entry point by collaboration stage, see the stage-to-skill table.
For open research, focused policy/market/competitive studies and product opportunity decisions, see the research Skills usage guide.
For project adoption of requirement/design/task constraints and the post-coding engineering + functional repair loop, see engineering quality governance.
- Four separated asset layers — Skill (what an agent can do) / Spec (what a thing looks like) / Protocol (how parties coordinate) / Rule (what must not happen), with explicit boundaries. See terminology.
- Orchestrators are thin — an
orchestrate-*skill does exactly four things: detect context, sequence calls, halt on failure, aggregate output. Domain logic stays in the atomic skills. - Review criteria live outside the skills — evaluative skills read their criteria from
rules/*-quality.md, so one definition serves the producing, diagnosing and reviewing sides. - Vendored-only distribution — externally derived skills are pinned to a commit and digest with their license recorded. Nothing is installed from the network at runtime. See ADR 0011.
mkdir -p ~/.local/share
git clone --depth 1 https://github.com/nesnilnehc/ai-cortex.git ~/.local/share/ai-cortex
~/.local/share/ai-cortex/bin/cortex installcortex install also puts the command itself at ~/.local/bin/cortex, which is where the short cortex used below comes from; it warns if ~/.local/bin is not on your PATH, and until it is, call the script by its full path.
It symlinks every skill — including reviewed local copies of externally derived ones — into ~/.agents/skills/<skill>, where Codex and other agents reading that path discover them in a new session. It also detects installed IDEs (Claude Code, Cursor) and syncs their skill paths. User-scoped Rules are symlinked for Claude Code and converted to .mdc for Cursor; project-scoped engineering Rule sets stay in the canonical clone and are loaded on demand by their review Skills, avoiding permanent context inflation. specs/ and protocols/ need no installation — agents read them from the canonical path. Nothing is ever installed from skills.sh or GitHub at runtime.
cortex updateFetches origin and hard-resets the canonical clone onto it, then re-syncs, pruning orphaned links for deleted skills and rules. It refuses to run on a dirty working tree; cortex update --force overrides that and discards whatever was uncommitted. Local commits on the tracked branch are discarded either way, so keep your own changes on a branch or a fork rather than in the canonical clone.
Upstream updates to externally derived skills are reviewed by maintainers before entering AI Cortex; they are not upgraded independently at the user's runtime.
cortex statusShows CORTEX_HOME, the current commit, link counts per IDE, and any legacy artifacts detected.
If this machine previously installed AI Cortex some other way, review before removing:
cortex clean --dry-run # report only
cortex clean # interactive, confirms each categorycortex uninstall # remove cortex-managed symlinks and .mdc files, keep CORTEX_HOME
cortex uninstall --remove-home # also delete the CORTEX_HOME directoryInstall design is recorded in ADR 0010; external skill management in ADR 0011. Engineering Rule profiles and review-gate responsibilities are recorded in ADR 0012.
See CONTRIBUTING.md, and read the Code of Conduct before taking part. Report security issues privately per the security policy rather than opening a public issue.
Maintainers releasing AI Cortex: docs/guides/releasing.md covers what counts as worth releasing, the independent version domains, and the Skills that own each step.
Documentation is written in English; the exceptions are listed in docs/LANGUAGE_SCHEME.md.
Original AI Cortex content is MIT. Vendored externally derived skills keep their own licenses — see the license policy and third-party notices.
- Contributors
- Pinned sources and local modifications for externally derived skills are listed in ATTRIBUTIONS.md and skills/SOURCES.yaml