跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

trip-planner

End-to-end international trip planning: turns "I want to go to X for N days" into a verified, bookable plan — route skeleton across cities, flight price scans (international + domestic legs), train-vs-fly decisions, hour-by-hour daily timelines with opening hours, dwell times, holiday collisions and tappable turn-by-turn map links (小时级行程+地图导航+离线KML), hotel shortlists by neighborhood, budget rollup, and a booking checklist with deep links. Use this whenever the user asks to plan a trip, vacation, itinerary or honeymoon, compare flight dates/prices, pick between cities or routes, schedule a travel day hour by hour, fill a spare block of time ("I'm near X with 2 free hours"), turn a finished plan into a designed page (eight themed renders: illustrated / clay / noir / glass / journal / zine / splash / portal — 插画/黏土/夜航/玻璃/手账/Zine/闪屏/穿越版), or asks 旅行规划/行程安排/机票比价/去某国玩N天怎么安排/现在有空档干嘛/把行程做成好看的网页 — even if they only mention one piece (just flights, just hotels, or just navigation), the playbook

skywain · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add skywain/trip-planner-skill --skill trip-planner

技能说明


name: trip-planner description: >- End-to-end international trip planning: turns "I want to go to X for N days" into a verified, bookable plan — route skeleton across cities, flight price scans (international + domestic legs), train-vs-fly decisions, hour-by-hour daily timelines with opening hours, dwell times, holiday collisions and tappable turn-by-turn map links (小时级行程+地图导航+离线KML), hotel shortlists by neighborhood, budget rollup, and a booking checklist with deep links. Use this whenever the user asks to plan a trip, vacation, itinerary or honeymoon, compare flight dates/prices, pick between cities or routes, schedule a travel day hour by hour, fill a spare block of time ("I'm near X with 2 free hours"), turn a finished plan into a designed page (eight themed renders: illustrated / clay / noir / glass / journal / zine / splash / portal — 插画/黏土/夜航/玻璃/手账/Zine/闪屏/穿越版), or asks 旅行规划/行程安排/机票比价/去某国玩N天怎么安排/现在有空档干嘛/把行程做成好看的网页 — even if they only mention one piece (just flights, just hotels, or just navigation), the playbook and verification rules here still apply.

Trip Planner

Turn a fuzzy trip idea into a plan the user can book link-by-link. The deliverable is verified and bookable, not inspirational: every price and opening time carries a source + as-of date, or an explicit "verify at link" flag. AI travel tools fail on stale data, not on prose — fixing exactly that is this skill's job, so verification IS the work.

Hard rules

  1. Never book, pay, hold, or enter personal data anywhere. Produce deep links and a checklist; the human books. This is what keeps the skill safe to run autonomously.
  2. Prices and hours come from tools, never from memory. Model memory is fine for geography and "what's worth seeing"; anything bookable or closable gets checked. A missing price is written "—, check link", never guessed.
  3. Cheap before expensive: bundled script + keyless APIs first (see references/data-sources.md), browser automation second and only for what scripts can't get (OTA hotel prices, LCC fares, odd venues). Never curl OTA/airline sites — they bot-block instantly; browser pane only. Pace requests like one polite human.
  4. Search budgets are real: ~25 web searches for your own orchestration work (visa, flights, holidays, hotels, assembly) — separate from, not inclusive of, the ≤8 written into each parallel city subagent's prompt. Unbounded research agents hang and burn money, so the cap goes in the prompt every time. Budget exhausted → ship with the least-verified items flagged rather than digging further.
  5. Reply in the language the user asked in. Report money in the user's home currency (infer from origin), stating the FX rate + date used once. FX source: frankfurter.dev first — but it only carries ~30 major currencies, and closed or minor ones (MAD / VND / EGP …) are not "unsupported", they are silently dropped from a 200 response (symbols=VND,USD comes back with USD alone). For those use https://open.er-api.com/v6/latest/<BASE> and check the returned object has the key you asked for; the plan states which source it used (data-sources.md §FX).
  6. Track the phases as todos (whatever task/todo tool the harness has; none → a short checklist at the top of your working notes) so a long plan survives interruptions and stays visible.

Interaction contract

Three moments at most, usually two: (0) one intake message, only if a core fact is missing and can't be inferred (Phase 0 — most requests need none); (a) after Phase 2 — present 2-3 route skeletons, get a pick; (b) final delivery. Everything else runs without questions. If the user says "一次到位 / don't ask, just plan" or the session is clearly headless, skip (0) and (a): assume, pick the best skeleton yourself and state every assumption prominently at the top of the output.

Quick modes (no full pipeline)

  • Gap filler — "I'm near X with 2 free hours": offer 2-3 options within a 15-min radius, one per energy level (a sight / food / a sit-down), each with walk time, a map link, a turn-back deadline, and — the one thing worth a search — confirmation that it is open right now. ≤3 searches; answer in minutes, not a report.
  • Single day — "we have one day in Rome, what do we do?": run Phase 1's holiday + festival check, Phase 4 for that one day, and the Phase 6 self-check. Skip route skeletons, flights and hotels entirely; read scheduling.md and navigation.md and leave the rest closed. This is the most common request that is not a whole trip.
  • Live replan — "missed the train / it's pouring": rebuild only the affected day from its degradation tags ([skippable]/[swap→…]/late_cut line) instead of re-planning the trip. [pinned] blocks hold; [opener] may move but costs a queue; re-verify only the hops that changed.

Phase 0 — Intake (one message, or none)

Read references/phase-0-intake.md now, before you decide whether to ask the user anything — it is the whole procedure for this phase (what counts as a core fact, how origin is inferred, the intake message format and its rules, what goes into prefs, the picture-capability check, the style line and the plan language, the exit criteria). This section is only the contract; do not compose an intake message from memory of the format.

Inputs: the user's request and anything said earlier. Outputs: the plan's top-level prefs block and lang, prefs.pictures (native | key | stock), the assumptions block for checkpoint (a) — and at most one intake message.

Gates — these decide pass/fail and do not move into the reference file:

  • One message, or none. Ask only when a core fact (origin · destination · when / how long · page style) is missing and cannot be inferred; ask for everything in ONE message in the intake format (core first, optional after, each optional line with its default, one "all defaults" line); an
阅读完整来源 ↗