Create and use skills
Skills package instructions, tools, and operating rules for repeatable agent work. Use them when a workflow or agent role needs a consistent way to perform a specialized task.
When To Create A Skill
Create a skill when you need:
- repeatable instructions for a role;
- a standard workflow for a domain;
- clear tool rules;
- reusable review criteria;
- consistent output expectations.
Do not create a skill for one-off instructions that belong in a workflow node prompt.
Skill Shape
A skill normally includes:
- name;
- description;
- when to use it;
- operating rules;
- tool guidance;
- output expectations;
- examples or templates.
In the repository, built-in skills live as src/skills/manifests/*.skill.json manifests. They are treated as operator-editable runtime data loaded from AGENT_HARNESS_SKILL_DIR. The TypeScript loader and import cache live in src/skills/index.ts; executable tools live separately under src/tools/.
Create A Skill
- Open the skills or capabilities settings area.
- Start a new skill.
- Write a short description that says when the skill should be used.
- Add operating rules.
- Add tool guidance and boundaries.
- Add expected outputs.
- Validate the skill.
- Import or save it.
- Attach it to an agent profile or workflow node.
- Run a small test.
Use A Skill In A Workflow
- Open the workflow designer.
- Select the agent node.
- Open agent settings.
- Choose the relevant skill.
- Confirm allowed tools and approval policy still match the skill.
- Save and test the workflow.
Review Skills Before Activation
Check that the skill:
- does not grant broader tool access than needed;
- states when it should not be used;
- keeps approval boundaries clear;
- produces inspectable artifacts;
- does not hide risky actions inside vague instructions.
Built-In Skills To Know
Current built-in skills cover repository mapping and editing, runtime verification, test triage, dashboard QA, UX usability audits, automation authoring, coordinator automation operations, repo-grounded research, model/provider evaluation, and bounded security workflows from attack-surface mapping through remediation retest and pentest reporting.
For the current list, see Capability Inventory.