Skills overview
agent-os ships its skills in three buckets. Most workflows are invoked by hand; check-work is an automatic review workflow. Disciplines trigger themselves from the situation, and the meta-skill governs how the others are written. Read this page top to bottom: the planning foundation, the remaining workflows, and finally the component skills the workflows compose.
The foundation: plan and shape
Planning starts here, and each layer builds on the one before it. plan-work owns the coherent mission and chooses whether a simple contract, repository coverage, focused questions, or a decision map is needed. shape-work materializes coherent delivery units and implementation-ready issues.
| Skill | Bucket | Invocation | Purpose |
|---|---|---|---|
plan-work | workflow | manual | Plan a coherent mission with adaptive depth and canonical decisions |
shape-work | workflow | manual | Shape coherent delivery units and implementation-ready issues |
Execution and operations
The workflows that run, pick, set up, and remember work.
| Skill | Bucket | Invocation | Purpose |
|---|---|---|---|
deliver-work | workflow | manual | Implement one change against boundaries and ground truth |
batch-work | workflow | manual | Run isolated ready units and verify the integrated result |
dispatch-next | workflow | manual | Pick or dispatch one action according to the request |
init-agent-os | workflow | manual | Managed policy setup or repository defaults |
record-lesson | workflow | manual | Record a durable lesson in repo or global policy |
simplifier | workflow | manual | Remove unnecessary code and solution layers |
simplifier-audit | workflow | manual | Audit a repository for simplification opportunities |
check-work | workflow | automatic | Independently review a candidate and report or fix supported findings |
Component skills
Standalone pieces that plan-work composes when useful. Invoke them directly whenever the piece is useful on its own — a grilling without the routing, or a plain-language summary of any plan, diff, or pull request.
| Skill | Bucket | Invocation | Purpose |
|---|---|---|---|
understand-work | workflow | manual | Question out the need behind a stated wish |
explain-work | workflow | manual | Explain the task in plain language for approval |
Disciplines
Always on; no invocation.
| Skill | Bucket | Invocation | Purpose |
|---|---|---|---|
verify-before-done | discipline | automatic | Fresh evidence before any completion claim |
diagnose-before-fix | discipline | automatic | Reproduce and root-cause before patching |
proportional-testing | discipline | automatic | Minimum meaningful regression coverage |
scope-guard | discipline | automatic | Keep work inside the task; flag drift |
simplifier-review | discipline | automatic | Review a diff for unnecessary complexity |
quality-ratchet | discipline | automatic | Compare exact entry and candidate evidence without score gates |
notice-lesson | discipline | automatic | Treat interruptions as misunderstanding signals |
plain-voice | discipline | automatic | Apply plain voice to every agent-authored text response |
Meta
| Skill | Bucket | Invocation | Purpose |
|---|---|---|---|
list-skills | meta | manual | List installed skills and how to invoke them |
writing-skills | meta | manual | Doctrine and definition of done for agent-os skills |
Invocation
Manual skills are opted into explicitly on both platforms. In Claude Code they carry disable-model-invocation: true in their frontmatter and are typed as /<skill> when installed directly or /agent-os:<skill> through the plugin. In Codex they carry an agents/openai.yaml with policy.allow_implicit_invocation: false and are typed as $<skill>.
Automatic skills carry neither switch. Their descriptions name the situation and boundary clearly enough to trigger only when useful; check-work may also be explicitly named with report or fix.
Retirement
A retired skill moves to the repository's root deprecated/ folder rather than being deleted. That folder sits outside the plugin's skills/ directory, so its contents are never distributed, and a skill re-enters skills/ only by passing the definition of done again.