Using uni admission agent
Skill dlfkid/uni-admission-agent/skills/using-uni-admission-agent
Entry point for the uni-admission-agent plugin. Use when the user wants to crawl, scrape, extract, fetch, install, diagnose, or export university admission / program / course / tuition / deadline data — or any operation involving the adm-agent tool. Triggers on phrases like "抓取大学课程", "爬取招生数据", "crawl university programs", "install adm-agent", "导出 Excel", "上次爬取失败".From its SKILL.md
npx -y skills add dlfkid/uni-admission-agent --skill using-uni-admission-agentAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.3 KB, as published. Nobody here has run it
uni-admission-agent — Router Skill
You are operating the adm-agent crawler tool (https://github.com/dlfkid/uni-admission-agent). It extracts structured admission / program / course / tuition / deadline data from university websites, with LLM-assisted parsing. Results land in a local SQLite database by default (Postgres optional).
This is the router skill. It does two things:
- Preflight — figure out whether the tool is even installed and running.
- Route — pick the right sub-skill based on the user's intent and current system state.
Always start here. Never jump straight into a sub-skill without checking preflight first — half of "the crawler isn't working" reports turn out to be "the server isn't running".
Preflight (run silently before routing)
# 1. Is the CLI installed?
command -v adm-agent >/dev/null && echo "cli=ok" || echo "cli=missing"
# 2. Is the server running?
curl -sS --max-time 3 http://127.0.0.1:8910/health >/dev/null && echo "server=ok" || echo "server=down"
Outcomes:
cli | server | Action |
|---|---|---|
| missing | * | Route to [[uni-admission-install]] — the tool isn't installed. |
| ok | down | Tell the user: "后端没启动,我帮你拉起来吗?" If yes → route to [[uni-admission-install]] §"Start an existing install". |
| ok | ok | Proceed to Routing below. |
Do NOT run adm-agent up or any server-starting command in the background yourself — it's a foreground process. Always defer to [[uni-admission-install]] for lifecycle ops.
Routing
Once preflight passes, pick the sub-skill that matches the user's intent:
| User intent (examples) | Sub-skill |
|---|---|
| "抓取 / 爬取 / extract / crawl" + university name or URL | [[uni-admission-crawl]] |
| "上次爬失败了 / quarantine / 数据质量 / 检查失败 / why did it fail" | [[uni-admission-diagnose]] |
| "删除 / 清空 <university> 的数据 / delete programs for" | [[uni-admission-diagnose]] (Step 5 — cleanup commands) |
| "导出 / Excel / CSV / 下载数据 / export" | [[uni-admission-export]] |
| "安装 / 升级 / 启动 / 重装 / install / upgrade / restart" | [[uni-admission-install]] |
| Mixed: "帮我爬利兹大学" but cli=missing | Install first, then auto-continue to crawl |
When the request mixes phases, do them sequentially in the same conversation — don't ask the user "ok do you want to crawl now?" after install completes. Just do it.
When the request is ambiguous (e.g., "看看这个工具能干嘛"), respond directly with a short menu — don't load any sub-skill yet.
Shared glossary
These terms appear across all sub-skills. Internalize them so you don't have to re-explain:
- slug — short URL-safe university identifier, regex
^[a-z0-9-]+$. Examples:hku,leeds,manchester. Never use display names like "The University of Hong Kong" as slug. - index page — a page that lists multiple programs (course catalog, search results).
- detail page — a page for one specific program (e.g., "MSc Finance" entry page).
- paginated crawl — multi-page index walk; auto-stops via signals (
url_drift,decreasing_yield,quality_failed). - quarantine — extracted records that failed quality gate. Visible via
adm-agent quarantine list. Not deleted; kept for review. - audit funnel — per-crawl counts at each pipeline stage: raw → filtered → candidates → extracted. Visible via
adm-agent audit list;adm-agent audit drill <ID>(an id from that list, not a university/year filter) goes one level deeper into the actual dropped URLs for one row. - stop_reason — why a paginated crawl stopped. Values:
exhausted(normal),max_pages(cap hit),url_drift(URL pattern broke — ⚠️),decreasing_yield(last few pages near-empty — ⚠️),quality_failed(quality gate blocked — ⚠️).
Conventions
- Never invent URLs. The user gives you the entry URL; you do not guess "this looks like a UCL program page so I'll try X".
- Never escalate to sudo / system install without explicit user confirmation. Even then, [[uni-admission-install]] §"manual install" must be followed verbatim.
- All long-running tasks need progress reporting. If you kick off a paginated crawl, poll
/tasks/<id>/eventsand stream meaningful events to the user (page progress, quality checks, stop signals) — don't just sleep silently. - Stop and ask when the user's request would cost > $0.50 in LLM calls or > 10 minutes of crawl time. Quote the estimate, get a yes.
When you arrive at a sub-skill
Each sub-skill assumes preflight passed and the user's intent is already classified. If you arrive at a sub-skill without going through this router, stop and load [[using-uni-admission-agent]] first — there's shared context (preflight state, glossary, conventions) that the sub-skill counts on you having read.