Skills

Skills

Skills are reusable workflows Warden can capture and re-invoke. Save a multi-step task once with create_skill; later, just say the example prompt and Warden activates it and reuses the workflow.

The lifecycle

create_skill (capture workflow) → activate_skill (load into context) → say the example prompt (Warden runs the workflow)

Deactivate with deactivate_skill. List everything with list_skills.

The commands

Command	What it does
create_skill	Captures a workflow: name, description, the example prompt that triggers it, and the steps Warden takes. Saved to the skills database.
list_skills	Returns every saved skill — name, description, example prompt, active state.
activate_skill	Loads a skill into Warden's active context so the next matching prompt reuses its workflow.
deactivate_skill	Removes a skill from active context. The saved skill still exists; it just won't auto-trigger.

Example: capturing a deploy-check workflow

Tell Warden what you want it to remember, with the trigger phrase:

Create a skill called "deploy-check" that I trigger with
"check the deploy". When I say that: SSH into the deploy host,
run `systemctl status dockbox`, check for errors in the last
100 log lines, and tell me if anything is failing.

Later, just say the trigger:

check the deploy

Warden activates the skill and runs the workflow — no need to repeat the steps.

When to make a skill

Make a skill when you find yourself giving Warden the same multi-step instruction more than once. Don't make skills for one-offs — that's just clutter in the trigger namespace.

Skills are best for: deploy checks, weekly reports, recurring audits, "look at this PR and tell me if X", any task with a fixed shape but variable inputs. The orchestrator decides when to activate a skill based on your prompt matching its example_prompt or when_to_use description.

Skills vs. agent delegation

Skills run inside the orchestrator loop — they parameterize what Warden is trying to do, but they are not new agents. If a skill needs code changes, file edits, or a deep technical implementation, the orchestrator will still delegate that work to Atlas or another agent while using the skill to shape the overall plan.