How the five Archgate skills — reviewer, lessons-learned, adr-author, onboard, and cli-reference — orchestrate the governance loop inside your AI coding agent.
A skill is a focused, reusable capability that your AI coding agent invokes at a specific point in the development workflow. Archgate ships five skills as part of the editor plugins, and together they turn the ADRs stored in .archgate/ into a closed governance loop: decisions guide code before it is written, code is checked against them after, and new lessons feed back into the decisions.
Skills are normally invoked by the developer agent automatically — you rarely call them by hand. Each one owns a piece of the workflow:
UNDERSTAND → PLAN → WRITE → VALIDATE → CAPTURE reviewer lessons-learnedonboard runs once at the start of a project to create the initial ADRs. adr-author is the authoritative authoring layer that every other skill delegates to whenever an ADR must be created or edited. cli-reference is an internal helper that loads canonical CLI documentation on demand.
The five skills
Section titled “The five skills”reviewer — the VALIDATE step
Section titled “reviewer — the VALIDATE step”After code is written, the reviewer skill validates it against the project’s ADRs. It gathers everything in one call with archgate review-context --run-checks (changed files grouped by domain, the applicable ADR briefings, and automated check results), then launches a focused sub-agent per affected domain to verify each one in isolation. If the automated checks already fail, it blocks immediately. ADR violations are hard blockers; warnings are always surfaced to the user but do not block. Running each domain in its own sub-agent keeps the parent agent’s context lean even when many files changed.
lessons-learned — the CAPTURE step
Section titled “lessons-learned — the CAPTURE step”The lessons-learned skill is the last step of the workflow, after checks and the reviewer confirm compliance. It reads the full session transcript via archgate session-context (recovering context that compaction may have dropped) and the current governance landscape via archgate review-context, identifies patterns and recurring mistakes, and proposes new ADRs or extensions to existing ones — or records operational notes in project memory. It is deliberately conservative: it only proposes rules for clear, repeated patterns backed by concrete code evidence. This is the learning movement of the loop that turns mistakes into permanent rules.
adr-author — the single authoring layer
Section titled “adr-author — the single authoring layer”The adr-author skill creates and edits ADRs to match the structure, tone, and depth of the project’s existing decisions. It enforces the six required sections, the frontmatter field order, the domain ID prefixes, and the companion rules file when automated enforcement is warranted. It is the only place that writes ADR files — onboard and lessons-learned delegate to it rather than touching ADRs directly, which keeps every decision in the project at a consistent standard.
onboard — first-time setup
Section titled “onboard — first-time setup”The onboard skill bootstraps a project into governance: it verifies the CLI, runs archgate init, explores the codebase with parallel sub-agents, interviews you about your conventions, proposes an ADR backlog, and authors every approved ADR in parallel before delivering a governance summary. Run it once when setting up Archgate, and again later to extend governance into new domains or audit existing ADRs.
cli-reference — internal reference loader
Section titled “cli-reference — internal reference loader”The cli-reference skill is invoked by other skills and the agent, not by you. Because the archgate CLI is newer than most model training cutoffs, agents must not guess at command syntax or the rule API. This skill fetches the canonical CLI documentation and returns only the section the caller asked for (CLI commands, rules authoring, or ADR format), so agents work from authoritative reference rather than memory.
How the skills compose the loop
Section titled “How the skills compose the loop”The skills map directly onto the three movements of the governance loop:
| Movement | Skill(s) | What happens |
|---|---|---|
| Prevention | onboard, adr-author | ADRs are created and kept current; the agent reads them |
| Detection | reviewer | Code is validated against ADRs after it is written |
| Learning | lessons-learned → adr-author | New patterns are captured back into ADRs |
cli-reference is the connective tissue: any skill that needs exact command or API detail loads it first.