ss-spec
GitHub用于启动非平凡功能开发的需求规格制定技能。通过捕获功能的WHAT和WHY,生成结构化的需求文档,并执行前提检查,在实施前确保需求清晰且可验证,等待人工审批后进入规划阶段。
触发场景
安装
npx skills add Serial-Studio/Serial-Studio --skill ss-spec -g -y
SKILL.md
Frontmatter
{
"name": "ss-spec",
"description": "Phase 1 of Serial Studio's spec-driven workflow: capture WHAT a feature must do and WHY, with no implementation detail. Use to start any non-trivial or multi-file feature — \"spec this out\", \"write a spec\", \"start a feature\", \"let's spec-drive X\". Produces doc\/claude\/specs\/NNNN-slug\/spec.md and gates on human approval before \/ss-plan.",
"argument-hint": "[feature name or one-line description]"
}
Serial Studio — /ss-spec (phase 1 of 4)
Capture the WHAT and the WHY. No file paths, class names, signal wiring, or algorithm
choices — those belong to /ss-plan. A spec that leaks implementation pre-commits the design
before it has been reviewed. See doc/claude/spec-driven.md
for the full workflow and the gate discipline.
Procedure
-
Pick the number and slug. The next number is
max(existing) + 1, zero-padded to four digits:ls -d doc/claude/specs/[0-9][0-9][0-9][0-9]-* 2>/dev/null \ | sed -E 's#.*/([0-9]{4}).*#\1#' | sort -n | tail -1If that prints nothing, start at
0001. Choose a short kebab-case slug from the feature name (0007-canbus-filter). -
Ground yourself first. Skim the relevant
doc/claude/sub-doc(s) for the area so the spec's claims about current behavior are true — but keep the spec implementation-free. Ground in real behavior (a report, a screenshot, a measured limit), not a hypothesis. Then state the problem and the one deciding constraint in a single chat sentence before drafting sections — a problem you can name is one the requirements will actually serve (doc/claude/j-space.md, verbalize-to-load). -
Create the directory and copy the template:
mkdir -p doc/claude/specs/NNNN-slug cp doc/claude/specs/templates/spec.md doc/claude/specs/NNNN-slug/spec.mdFill in the frontmatter (
spec,title,created,author: Alex Spataru,status: draft). -
Write the spec into every section: Problem/Motivation, Goals, Non-Goals, Requirements (numbered + testable), Acceptance Criteria (tied to a real check), Constraints & Invariants, Open Questions.
-
Run the premise check. Name the one measurement that would kill the spec (the coupling figure that retired spec 0069, the local benchmark that would have stopped spec 0090 before eleven commits), run it time-boxed to about an hour in the scratchpad or against the running app, and record the question, the number and the verdict in the spec's Premise Check section. If the premise cannot be measured, say so there in one sentence; the maintainer decides whether to proceed on reasoning alone.
-
Resolve ambiguity before finishing. Where a requirement is genuinely the maintainer's call (scope, tier gating, UX shape), use
AskUserQuestionwith a recommendation — do not guess and bury the assumption. Unresolved items stay in Open Questions.
Gate
Stop after writing spec.md. Present it for review and do not run /ss-plan until the
maintainer approves it (set status: approved). The spec is the contract every later phase
is measured against; an unreviewed spec poisons the plan and the build.
Rules
- WHAT/WHY only. The first reviewer should be able to read it without seeing the codebase.
- Acceptance criteria must be verifiable — prefer ones that map to
pytest,tests/scripts/,--benchmark-hotpath, or a concrete in-app observation. - This file is the only thing you create. No code, no plan, no scaffolding yet.
- The Premise Check section never keeps its template text:
/ss-planrefuses a spec without a recorded measurement or an explicit one-sentence "unmeasurable" verdict. - When the maintainer asks for a commit, the spec directory goes in its own
docs(spec):commit before any implementation;scripts/commit-shape.pyrefuses the mix.
版本历史
-
3cbdc24
当前 2026-11-03 11:14
新增 Premise Check(前提检查)部分到规范模板,要求在执行计划前进行测量验证;引入 commit-shape gate 和 mistakes-ledger ratchet 等配套工具以强化流程纪律。
- cb2f5d2 2026-08-20 08:59


