From dev
Implement the next pending task from tasks.yml. One task per invocation — user loops externally. Triggers on: "dev/build", "build next task", "implement next"
How this skill is triggered — by the user, by Claude, or both
Slash command
/dev:buildThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Implements exactly ONE task per invocation. The user runs this in a loop externally — each call gets a fresh context window.
Implements exactly ONE task per invocation. The user runs this in a loop externally — each call gets a fresh context window.
Read .dev/tasks.yml and .dev/progress.md (if it exists).
If .dev/tasks.yml doesn't exist, stop and tell the user to run dev/how first.
If .dev/progress.md exists, read the Codebase Patterns section at the top — these are consolidated learnings from previous iterations.
Run each requirement's check command from .dev/tasks.yml. If any fail, report the failures and stop — status: blocked.
First, check if any task has status: blocked or status: fatal. If so, stop immediately — report the blocking task and its notes. The user must resolve it before the loop continues.
Then select the first task with status: pending. If no pending tasks remain, report completion and stop.
Update the task's status to in_progress in .dev/tasks.yml.
Implement the task according to its description, respecting:
files section (create or modify)If you discover the task is impossible or the plan is wrong, do not hack around it. Set status to fatal in .dev/tasks.yml, explain why in the task's notes, and stop. The user must go back to dev/what or dev/how to fix the plan.
Run these in order. All must pass before proceeding.
If checks fail, fix the issues. Maximum 3 attempts per check type — if still failing after 3 attempts, set status to blocked, explain in notes, and stop.
Run code-review with a prompt describing what was implemented and why:
code-review "Implemented [task title]: [brief description of what was done and the intent behind it, referencing task ID from tasks.yml]"
code-review is a long-running CLI — use an extended timeout.
If review returns actionable feedback, address it and re-run. Maximum 3 review cycles — if still getting material feedback, address what you can and note remaining items.
See references/build-protocol.md for the full code review integration.
done in .dev/tasks.ymlnotes field with anything the next iteration should know.dev/progress.md (see references/build-protocol.md for format)Stage all changes — implementation files, .dev/tasks.yml, and .dev/progress.md — in a single commit:
feat: [task-id] - [task title]
The tree must be clean after this step.
End your response with a clear signal:
status: donedev/what or dev/howfiles list without updating .dev/tasks.yml firstfatal — don't improvise.dev/progress.md's Codebase Patterns section before implementingGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub codethread/claude-code-plugins --plugin dev