Agent Skills › Serial-Studio/Serial-Studio › ss-tasks

ss-tasks

GitHub

将已批准的计划分解为有序、可独立验证的任务清单。确保任务覆盖所有文件及验收标准，明确依赖与验证方式，并在实施前等待人工审批，以保障开发流程的清晰与可控。

.claude/skills/ss-tasks/SKILL.md Serial-Studio/Serial-Studio

Trigger Scenarios

/ss-plan is approved break this into tasks task it out

Install

npx skills add Serial-Studio/Serial-Studio --skill ss-tasks -g -y

More Options

Non-standard path

npx skills add https://github.com/Serial-Studio/Serial-Studio/tree/master/.claude/skills/ss-tasks -g -y

Use without installing

npx skills use Serial-Studio/Serial-Studio@ss-tasks

指定 Agent (Claude Code)

npx skills add Serial-Studio/Serial-Studio --skill ss-tasks -a claude-code -g -y

安装 repo 全部 skill

npx skills add Serial-Studio/Serial-Studio --all -g -y

预览 repo 内 skill

npx skills add Serial-Studio/Serial-Studio --list

SKILL.md

Frontmatter

{
    "name": "ss-tasks",
    "description": "Phase 3 of Serial Studio's spec-driven workflow: decompose an approved plan.md into an ordered, individually-verifiable task checklist (tasks.md). Use after \/ss-plan is approved, or \"break this into tasks\", \"task it out\". Gates on human approval before \/ss-implement.",
    "argument-hint": "[spec number or slug]"
}

Serial Studio — /ss-tasks (phase 3 of 4)

Decompose the approved plan.md into an ordered checklist of small, individually-verifiable units. Each task is a coherent diff a reviewer could read on its own. See doc/claude/spec-driven.md for the gate discipline.

Preconditions

doc/claude/specs/NNNN-slug/plan.md exists and its status: is approved.

Procedure

Read spec.md and plan.md. The tasks must cover every file in the plan's affected- files table and every acceptance criterion in the spec — nothing more, nothing less.

Copy the template and write the tasks:

cp doc/claude/specs/templates/tasks.md doc/claude/specs/NNNN-slug/tasks.md

Decompose well:
- One task = one focused change. If a task touches more than ~3 files or needs a paragraph to describe, split it.
- Order so the tree stays coherent: data model / Keys:: and enums before the code that reads them; a driver class before its registration touch-points; UI after the model that feeds it.
- Each task names its Files, what it Does, how to Verify it (usually python scripts/code-verify.py --check <files> plus a test or read-back), and its Deps.
- A task touching a silent-breakage class (hotpath, ctor closure, signal wiring, cached hotpath flags) names the binding invariant in its Does line — so /ss-implement re-states it at edit time instead of rediscovering it (doc/claude/j-space.md: constraints steer only when named close to the action).
- Sequence hotpath-sensitive work so --benchmark-hotpath can be run at a clean point.
Fill the Definition of Done — the whole-feature gate (code-verify clean, qt-cpp-review run, hotpath not regressed, pytest targets identified, sanitize-commit run, diff in scope, spec marked done).

Gate

Stop after writing tasks.md. Present the checklist for review. Do not run /ss-implement until the maintainer approves it. Reordering or splitting tasks now is free; discovering a missing task mid-build is not.

Rules

Tasks describe work; they do not perform it. No code in this phase.
Every acceptance criterion in spec.md must be reachable by completing the listed tasks. If one is not, the plan is incomplete — go back to /ss-plan.
Keep the checklist honest: it becomes the live progress record that /ss-implement updates.