← ClaudeAtlas

buildlisted

Implement the pending tasks from a plan doc — each verified and patch-snapshotted, run to completion or a judgment call, left in the working tree. Never commits.
ledtorch/coding-agent-workflow · ★ 0 · Web & Frontend · score 70
Install: claude install-skill ledtorch/coding-agent-workflow
## Input The text after the command — `/build <plan-doc-path>` (Claude Code) or `$build <plan-doc-path>` (Codex) — optionally followed by a task number. Called **the plan doc path** below. ## Context (gather first) Before anything else, run these and read the output: ``` git branch --show-current git status --short ``` ## Charter **Implement the pending tasks from a plan doc, one at a time in order, each verified before the next starts. Run until the plan is done or a judgment call needs the user.** All work is left uncommitted for one holistic review before [`/commit`](../commit/SKILL.md). `/build` is the execution half of [`plan.md`](../plan/SKILL.md). Plan decides *what* and *in what order*; build executes. It never re-plans, never reorders, and never invents a task that isn't in the doc. > **Why batch is safe without commits.** Checkpointing (commits between tasks) and batching > (proceeding to the next task) are different things. Commits stay forbidden > (this workflow's standing rule 1, restated in `references/SKILL.md` where that exists); > what actually gates the run is **per-task verification** — a task that doesn't verify > clean halts the run before it contaminates later work — plus **per-task patch snapshots** > (below) as commit-free rollback points. > > **The honest trade:** without a user review between tasks, a slightly-off interpretation > in an early task can propagate before the user sees it. The per-task acceptance criteria > and *Files likely touc