A Claude Code skill that turns a plan discussion into a full spec, ready to hand off to an AI agent for implementation.
If you've used spec-driven development workflows before, this is that: a portable, self-contained Claude Code skill you can drop into any repo.
After you discuss a feature with Claude, run /spec feature-name and it writes five files:
plans/feature-name/product.md: the why, problem, goal, reasoning, scopeplans/feature-name/requirements.md: user stories, acceptance criteria, dependenciesplans/feature-name/design.md: file changes with line numbers, reuse audit, edge casesplans/feature-name/tasks.md: JSON task list with tests and verification commandsplans/feature-name/kickoff.md: a ready-to-paste implementation prompt
Then it commits the spec. When you run the kickoff prompt, the agent implements, tests, archives the spec to plans/completed/, and commits.
Here's what a kickoff.md looks like after running /spec rate-limiting:
Add per-user rate limiting to the API so abusive clients can't degrade service
for everyone else.
Read these files before starting:
- plans/rate-limiting/product.md
- plans/rate-limiting/requirements.md
- plans/rate-limiting/design.md
- plans/rate-limiting/tasks.md
Then implement each task in order. Rules:
- Complete tasks in the order listed in tasks.md
- Do not change any file not listed in the tasks
- After each task, run the task's verification command and confirm the expected outcome
- After all tasks, run the full relevant test suite and confirm it passes
- After the full test suite passes, move the entire plans/rate-limiting/ directory
into plans/completed/rate-limiting/
- Commit all changes with a descriptive message reflecting what was implemented.
This is the final step.
Clean enough to paste directly. The five spec files it references contain all the detail: file paths, line numbers, edge cases, and test scaffolds.
Copy the skill file into your project:
your-project/
.claude/
skills/
spec/
SKILL.md ← this file
That's it. Claude Code picks it up automatically.
- Use the Plan agent (
/planin Claude Code) to think through the approach first. It's better at architectural reasoning and will surface trade-offs before you commit to a design. - Once the plan looks right, run
/spec your-feature-namein the same conversation. - Claude writes the five spec files and commits them.
- Paste
plans/your-feature-name/kickoff.mdinto a new Claude session to implement.
- Claude Code
- A git repo (the skill commits the spec files)