菜单
Capture the headline ask — what field the user wants certified in, which capability areas the credential must cover, and what is out of scope. Writes profile/spec.md. First stage of the spec-led certification workflow; run it before any research.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/danielrosehill/spec-led-certification-plugin --skill define-spec命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Create a private GitHub repository from the Spec-Led-Certification template and clone it, as the durable home for one certification search. Use when the user wants the search backed by a repo rather than a local folder — "make a repo for this", "set up a workspace for finding a certification". For a plain local directory with no remote, use scaffold-workspace instead.
Create the directory structure a spec-led certification search writes into — profile/, scorecard/, research/, report/ — by copying the plugin's bundled template into the working directory. Idempotent and non-destructive: it never overwrites a file that already has content. Run this first when starting a search outside a clone of the Spec-Led-Certification template repo.
Second and later runs — archive the previous run, confirm the profile is still current, re-survey the market and report what changed since last time. Use when the intake is already on file and the user wants to know whether the market has moved, or whether an earlier recommendation still holds. For a first run use start-search; to change stored answers first use update-profile.
| name | define-spec |
| description | Capture the headline ask — what field the user wants certified in, which capability areas the credential must cover, and what is out of scope. Writes profile/spec.md. First stage of the spec-led certification workflow; run it before any research. |
The opening interview. It establishes what is being looked for, in enough detail
that a scorecard can later be built from it. It does not establish why, how much, or
by when — those are define-objectives and define-budget. Keep to scope here; the
modularity is deliberate, because on a re-run this file usually survives and the
others may not.
Read profile/spec.md. If status is not unfilled and review_after is in the
future, show the user what is on file and ask whether it still holds. A confirmation
is a one-line exchange and a bumped captured date — not a re-interview.
Everything learned goes into profile/ in this repository. Nothing goes into the
agent's own memory store. Not Claude's memory, not a hosted user profile, not
"I'll remember this". The repo is the memory, and it is the repo the user carries to
whatever assistant they use next. If you have a memory tool available, this is the
one case where you do not reach for it.
One question at a time. The user is thinking as they answer, and a wall of six questions gets six thin answers.
1. The subject. What field, in their own words. Let them ramble; the rambling contains the topic list.
2. The named topics. This is the part that matters and the part users skip. Push for concrete capability areas, then test each one: could I check a published exam blueprint and tell whether it covers this? "Agentic AI" fails that test. "Sub-agent delegation" passes it. Aim for four to ten.
If they cannot produce topics, offer a starting list drawn from the subject and ask them to cut and correct it — recognition is easier than recall. Mark anything you supplied and they merely accepted, because a topic the user did not actually ask for should not carry much weight later.
3. The shape. Proctored exam, assessed project, course completion, a portfolio review? One credential, or a backbone plus a specialist? Some users have no view; record that rather than inventing one.
4. Out of scope. What have they already ruled out, and why. This is worth its own question because it saves the research phase re-treading ground and stops a re-run six months later resurfacing the same rejected options.
Do not name candidate credentials. You almost certainly know some. Naming one here anchors the entire process on it, and everything downstream — the topic list, the weights, the research — bends toward the thing that was mentioned first. If the user names one, record it under open questions as a candidate to check, not as a topic.
Do not let the subject become the topic list. "I want an agentic AI cert" is a subject, and a scorecard built from it alone cannot distinguish two credentials.
Vague is fine, unresolved is not. If the user genuinely does not know something, write it under open questions. An honest gap is workable; an invented answer propagates into a weight and then into a ranking.
Replace profile/spec.md entirely, keeping the frontmatter shape:
status: filled
captured: <today, YYYY-MM-DD>
volatility: medium
review_after: <today + 12 months>
Then tell the user, in two lines, what you recorded and what remains open. Do not
proceed to the next stage without being asked — unless you are running under the
spec-led-certification meta-skill, which sequences the stages itself.