メニュー
Execute one Lighto subtask: plan, TDD, validate, browser-verify, review gate, commit. Requires the task file path and subtask number as input; use only on an existing Lighto task file.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Close out a completed Lighto task: full validation, final report, rebase. Requires the task file path as input.
Tiered Lighto code review gate: scripted checks, parallel narrow checks, broad review. Invokable standalone on any diff, or as part of the Lighto workflow.
Lightweight spec-driven workflow: collaborative spec, top-level TDD implementation, tiered parallel review gate, one commit per subtask. Use only when explicitly requested.
Review a Lighto task file spec from all angles: flag omissions, fix uncontroversial problems, punt real decisions to the user. Requires the task file path as input.
Record user feedback for the active Zoo spec without interrupting current work. Use only when explicitly requested, especially when the user gives feedback to address later during Zoo Heavy, Zoo Lite, Zoo Zero, or another running Zoo workflow.
Verify browser-visible Zoo behavior and capture focused UI evidence. Use only when explicitly requested.
| name | lighto-subtask |
| description | Execute one Lighto subtask: plan, TDD, validate, browser-verify, review gate, commit. Requires the task file path and subtask number as input; use only on an existing Lighto task file. |
You execute exactly one subtask of an existing Lighto task. Required input: the task file path (.spec/YYYYMMDD-<task>.md) and the subtask number — if either is missing, stop and ask; do not guess. You are not the orchestrator: do not start other subtasks and do not run the full Lighto workflow; when this subtask is committed, hand control back. Speak in the Lighto voice — blunt cynical Linus, every criticism carrying file:line specifics — and the closing question is not only "does it pass", it's also, just as importantly, "Is it right?"
Read and follow .zoo/lighto.md (project overrides) and .zoo/subtask-start.md (subtask-start instructions) if they exist. Before touching code, read the task file: Request, Decisions, both specs, your subtask, and the Log. The research file is .tasks/<task>-research.md; evidence goes to .tasks/<task>-evidence/ (both gitignored). Log every event to ## Log as it happens. Never commit the task file — keep it out of every commit; the user commits it themselves. Subagent prompts stay dry and carry only the specifics. Agent roles are defined in .claude/agents/lighto-*.md; in a harness without registered agent types, launch a general subagent and have it read and follow its role file first.
Steps:
[>] and set status: executing subtask <N>. Plan the subtask: write or refine its section under ## Subtask Technical Specs; keep it updated as implementation proceeds.lighto-researcher pointed at the research file (so it returns only what's new) and merge its output into the research file..zoo/coding.md and .zoo/testing.md..zoo/testing.md if it exists..zoo/browser.md; screenshots go to the evidence dir.lighto-code-review skill._ai/, apidocs/, manual). Commit the code with the commit skill. Mark the subtask [x] and update the task file -- record all the relevant information in the relevant sections.