Skip to content

skills.md

Latest commit

 

History

History
73 lines (58 loc) · 3.22 KB

skills.md

File metadata and controls

73 lines (58 loc) · 3.22 KB
Raw

Claude-compatible skills

Skills are on-demand instruction bundles the agent can load mid-session via the skill tool. dirge reads the same format as Claude and opencode.

Place skill directories in .claude/skills/, .opencode/skills/, .agents/skills/, or .dirge/skills/ in your project or home directory. Each skill is a directory containing SKILL.md with optional YAML frontmatter:

---
name: my-skill
description: A helpful skill
---
# Instructions
Detailed skill content here.

Skills are auto-discovered at agent startup and listed in the system prompt, ordered by salience (see below) so the most useful skills surface first. The agent calls skill "my-skill" to load the full content on demand. Project skills override global skills by name.

Learning a skill: /learn

/learn turns source material into a reusable skill without hand-authoring it:

/learn ~/projects/acme-sdk, focus on auth + pagination
/learn https://docs.example.com/api/quickstart
/learn filing an expense: open the portal, New > Expense, attach receipt, submit
/learn            # bare: distill the workflow from the current conversation

There's no separate distillation engine — /learn builds one standards-guided instruction that the agent runs with its normal tools (read, grep, find_files, webfetch), then saves the result through the skill tool. It gathers every source named, applies any focus/constraints in the request, and writes one SKILL.md.

Every created skill must include a ## Verification section — a single command that proves the skill works. Creation is rejected without it, and a freshly learned skill starts with one recorded success (it was validated in the session that produced it) so its salience begins grounded.

Salience and curation

Skill telemetry lives in the project's SQLite database (.dirge/sessions/state.db), reusing the same salience engine as memory:

  • Reinforce on use — loading a skill nudges its salience up.
  • Disuse decay — the background curator decays skills nobody consults.
  • Effectiveness — a skill's success/failure record folds into its ranking, so one that keeps working outranks one that has failed in practice.
  • Confidence — a tiebreak among equally-salient skills.

The curator archives an agent-created skill once its effective salience falls to the archival threshold — i.e. it's gone unused long enough that decay overtakes any track record. A skill that keeps proving useful survives on its effectiveness even if it hasn't been touched recently; pinned skills and skills you didn't author (bundled/manual) are never auto-archived. Archived skills move to .dirge/skills/.archive/ and are recoverable. This replaces the earlier age-only staleness rule so the library self-prunes on usefulness, not just recency.

Bundled starter skills

The repo ships a small pack of general-purpose workflow skills under skills/ — systematic-debugging, code-review-feedback, and writing-skills. They are not installed automatically; copy the ones you want into a discovered location, e.g. cp -r skills/systematic-debugging .dirge/skills/ (per project) or ~/.dirge/skills/ (global). See skills/README.md.