A tight, opinionated workflow library for Claude Code. Twelve skills covering the full loop from "I have an idea" to "shipped and verified."
/plugin marketplace add Coherence-Daddy/cd-skills
/plugin install cd@cd-skills
/reload-plugins
Invoke any skill via the Skill tool with name cd:<name>, or type /cd:<name> in chat.
| Phase | Skill | Purpose |
|---|---|---|
| Pre | cd:context |
Load enough context before acting. Memory + git + key files + a written "know / assume / verify" gate. |
| 1 | cd:brainstorm |
Refine intent into a written spec. Decomposes oversized ideas. |
| 2 | cd:plan |
Spec → checkbox plan. Approval gate before execute. |
| 3 | cd:execute |
Walk the plan step-by-step. Verification baked in. |
| 3a | cd:tdd |
Test-first mode of execute. |
| 3b | cd:parallel |
Independent steps → subagents. |
| 4 | cd:review |
Two modes: request and receive. |
| 5 | cd:ship |
Verify → commit → integrate. |
| X | cd:debug |
Root-cause-first debugging. |
| X | cd:worktree |
Physical isolation via git worktrees. |
| Meta | cd:new-skill |
Author new skills in CD format. |
| Meta | cd:memory |
Recall / save / audit harness auto-memory. Three modes. |
context → brainstorm → plan → execute → review → ship
└ tdd / parallel
└ debug (when stuck)
cd:context runs first when starting fresh in a repo. The phase skills are gates: each one ends by handing off to the next.
Most agent-skill libraries are noisy. CD is built on three rules:
- Each skill earns its slot. No skill duplicates another. No skill is a pep talk.
- Skills load on demand. Session-start injection is a 500-byte index, not a 5 KB manifesto.
- Gates over guidelines. Every skill ends with a hard handoff. Brainstorm hands to plan. Plan hands to execute. Execute hands to ship. The model can't drift past a gate without confirming.
MIT. Use, fork, and reshape it.