Skip to content
SkDD logo Skills-Driven Development

Skills-Driven Development

Agents that learn by doing — and remember how they did it.
Why SkDD? Quick start GitHub

The one-minute pitch

A skill is a reusable, discoverable playbook — a markdown file plus optional scripts — that an agent follows to accomplish a specific, repeatable task. SkDD adds a lifecycle on top of the Agent Skills spec: forge → register → discover → evolve → archive. The skills get better the more your agents work, because your agents are the ones forging and evolving them.

Quick start

Terminal window
pnpm dlx @zakelfassi/skdd init --harness=claude       # or codex, cursor, copilot, gemini, opencode, goose, amp

That one command scaffolds a canonical skills/ directory, writes .skills-registry.md, appends a Skills block to your harness instruction file, and materializes a .claude/skills symlink → ../skills so the harness finds the colony at its conventional path. See Configuration for details.

Explore

Why SkDD? The pitch, the lifecycle diagram, and who it's for. Start here if you want the narrative.
Configuration Per-harness wiring: Claude Code, Codex, Cursor, Copilot, Gemini CLI, OpenCode, Goose, Amp.
Skill colony concept What a colony is, how it differs from a static library, and the architecture diagram.
Forging skills The skillforge meta-skill checklist and what the agent actually does when you say 'forge this.'
Specification alignment How SkDD maps onto the agentskills.io v1 spec — what's spec-native vs SkDD-added.
Harness integrations Dedicated guides for 11 harnesses: Claude Code, Codex, Cursor, Copilot, Gemini CLI, OpenCode, Goose, Amp, VS Code, Junie, Roo Code.

Principles

  1. Forge, don’t front-load — let agents create skills when they notice patterns during real work.
  2. Small skills, composed loosely — complex workflows emerge from composing small skills.
  3. Skills are living documents — agents should evolve them when they hit edge cases.
  4. The colony is the product — invest in the registry and discovery mechanisms.
  5. Human-readable, machine-executable — markdown for humans, frontmatter for agents.