trip-planner
UnverifiedEnd-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.
references/scheduling.md
View on GitHub# Hour-level scheduling
Read this before assembling any day into a timeline. Hour-level timelines are the
default deliverable; deliver day-level granularity only when the user explicitly wants
a rough cut.
Contents: dwell times · assembly method · day types · traps · verification · format.
## Dwell-time defaults (first visit, tourist median)
| Venue type | Time |
|---|---|
| Flagship museum (Louvre/Vatican/British class) | 3-4 h |
| Standard museum / gallery | 2 h |
| Small museum / single exhibition | 1-1.5 h |
| Temple, shrine, church, mosque (a visit) | 30-45 min — active worship sites: see Traps |
| Headliner temple with an approach street (Kiyomizu, Kinkaku-ji, Todai-ji) | 1-1.5 h incl. the approach |
| Open-air archaeological site (Roman Forum, Pompeii, Ephesus, Acropolis) | 1.5-3 h, +30% if hilly and unshaded |
| Amphitheatre / arena (Colosseum, Arena di Verona) | 1-1.5 h |
| Major complex (Fushimi Inari full hike, Alhambra, Forbidden City) | 2-3 h |
| Castle / palace | 1.5-2.5 h |
| Observation deck / viewpoint | 45-60 min + queue (sunset slot: also the queue for the lift down, up to ~1 h) |
| Market / food street | 60-90 min |
| Historic district stroll | 1.5-2 h |
| Garden / park | 45-90 min |
| Aquarium / zoo | 2.5-3 h |
| Theme park | full day |
| Photo-stop landmark | 15-20 min |
These are defaults, not answers: whenever a venue is a headliner for this user,
research its realistic visit time and note the source. Scale by pace: relaxed ×1.3,
packed ×0.8, kids or mobility flags ×1.3.
## Assembly method (per day)
1. **Pin the immovables first**: timed-entry tickets, reserved meals/shows, intercity
departures, sunrise/sunset shots — these are `[pinned]`, contractually immovable.
The crowd-window opener from rule 2 is different: it is `[opener]` — movable, but
moving it costs an hour of queueing. Degradation and live-replan key off this
distinction, so keep the two tags apart. Everything else schedules around them.
Early-arrival margin is **tiered**, because a timed ticket buys a place in the
security line, not entry: 15 min at ordinary venues; **30-45 min wherever there is
security screening or an ID check** (Vatican, Eiffel, Sagrada Família, Forbidden
City, teamLab…). And never plan to buy at the door at a flagship — ticket-office
queues run 30-60+ min against near-zero online. If it can be bought online, even
same-day, that IS the plan; an unavoidable door purchase gets its own +45 min block
in the chain so the arithmetic sees it. **Activity operators publish their own
check-in lead** (rides, boats, helicopters, ziplines — often longer than the tier):
read it off the booking confirmation, where it replaces the tier, and take the
activity's real duration from the same page, not the product title (live case: a
ride sold as 1.5 h ran 2 h, and a "take the later slot if late" fallback would have
collided with the next check-in).
2. **Anchor the opener**: put the day's most crowd-sensitive venue at its opening
time. Top sites are empty in their first hour; the same visit at 11:00 costs
30-60 min of queueing. Crowd calendar beats the clock, though: put flagships on
weekdays and save parks/markets/neighborhood walks for the weekend; check each
headliner for **free-admission days** (commonly the first Sunday; Vatican Museums
the last Sunday) — the free day is the busiest day of the month, so avoid it or
treat opening-time arrival as mandatory; and the day after a venue's weekly closure
day (Tuesday after a Monday closure) carries spillover crowds. When the date is
fixed and lands on a weekend anyway — a layover, a city break, "we have Sunday
free" — the rule still has a move: give the flagship the day's **first** entry
slot, take the top of its arrival-margin tier (45 min), and name the weekend tax in
the plan instead of pretending it isn't there.
3. **Chain by geography** in the day's cluster order: compute every hop with
scripts/route_tools.py and verify the load-bearing hops per navigation.md. Each
block = arrival + dwell + buffer.
4. **Buffers are policy, not padding.** "Hop time" means **door to door**: access
walk + worst-case headway + in-vehicle + egress walk — and *then* the buffer
(+10 min per urban hop; +15 min per unfamiliar interchange; +30 min after any
luggage move). Counting only the ride is how a 25-minute block swallows a
29-minute journey. The corollary is worth knowing: under roughly 1.2 km in a city,
walking beats a one-stop metro once the wait is counted — so don't schedule the
ride. Jet lag is real even when the user says it isn't (see Day types).
5. **Meals**: lunch 11:30-13:30, dinner 18:00-20:30 (Spain: ~14:00/21:00 — check
country notes), placed in a food area within 10 min of the adjacent cluster,
60-90 min per meal.
6. **Energy curve**: at most 2 heavy anchors (>2 h) per day and never back-to-back,
**and never more than ~3.5 h of continuous on-feet anchor time without a sit-down
block** — the second test is the one that matters, because shaving two anchors
from 2 h to 1 h 45 satisfies the first while leaving four unbroken hours on stone.
One low-effort block (garden, café, shopping street) mid-afternoon — but see the
siesta trap before putting shopping there; cap time-on-feet at ~8 h (6 h with kids
or mobility flags).
7. **Golden hour**: run `python3 scripts/route_tools.py sun plan.geo.json --write`
once the days have a city-level coordinate (any stop's `lat/lon`, or the
Open-Meteo geocode from Phase 1 dropped into `stops[0]` — a venue 5 km away moves
sunset by well under a minute) — and run it **before you write a single sunrise /
golden-hour / dark-start sentence**, not after. The local clock is tzdata's, not
yours: Morocco leaves UTC+1 for UTC+0 on 2026-09-20, so the "07:15 sunrise" a
tester wrote from habit was an hour off on all ten days, and nothing downstream
(`check`, `qc.py`, the renderers) compares prose against `days[].sun`. Order:
coordinates → `sun --write` → read the values → write the prose.
It fetches civil dawn / sunrise / sunset per day
from sunrise-sunset.org (one request per day — keyed on **that day's first stop
with coordinates** + date + tz, cached next to the plan — **except on a moving
day, where it takes the day's LAST stop with coordinates**. A day is "moving"
when it carries `"travel_day": true` **or** its first and last stops with
coordinates are more than **150 km** apart; the script prints which rule fired
("last stop, first->last 1,090 km" / "(travel_day)"). Reason: the evening anchor
and the sunset that matters are where you sleep, not where you woke up — the
China test's Xi'an→Beijing day reported Xi'an's 17:41 against a Beijing anchor
at 16:59 sunset, 42 min wrong on exactly the day that squeezes in an evening
block. Order the moving day's `stops` in visit order with the arrival city last
and the rule does the right thing. **When the day's sun-critical anchor is at the
first stop instead** — a Chefchaouen sunrise before the flight to Casablanca, an
Erg Chebbi dune sunrise before the drive to Fes — the last-stop default is 25 min
wrong on exactly the block that cares, so override it: set the day's
`"sun_stop"` to that stop's `name` (or its 0-based index in `stops`) and `sun`
keys the day there. The header still prints one 天亮 for the whole day, so on a
long east–west move say in the day note that the other city's dawn/dusk
differs),
**sanity-checks the answer** (rejects `status≠OK`, two cities or two dates
returning identical times, and a day length that cannot belong to that latitude
and month — the failure mode that once returned equatorial 12 h 09 m for Norway
in October with `status: OK`), and writes each day's `sun` string in place.
**Read the tail of its output**: after the "N request(s), N day(s) written" line
it prints — and repeats as a `WARN` on stderr — `skipped/failed (N): <date>
(reason), …` naming **every day it did not write** — the reasons are: no ISO
date; no stop with coordinates; `tz approximate` (fix: set `days[].tz`, a
plan-level `tz`, or pass `--tz Area/City` — the longitude guess is wrong
wherever zones bend); request failed — plus one `sun REJECTED — …` WARN per
day the sanity checks refused. A run that says "6 written" on a 7-day plan now names
the missing date, so you re-run `--only DATE` instead of counting `sun` fields
by hand (Italy test F5) — and it **exits non-zero (1)** whenever that list is
non-empty, so a "9/10 written" run cannot pass unnoticed in a pipeline (Vietnam
test F6: one TLS failure, exit 0, nearly shipped). The written days are kept;
fix the named ones with `--only` and re-run until it exits 0. Exit **3** = a day
was REJECTED by the sanity checks (look before retrying — a "polar day/night"
rejection is not retried: remove that day's `sun` key (absent — not `null`, not
`""`), output-template.md §`sun`, PLN-11). A day with **no stops
at all** (pure travel/rest day) is informational only here — not counted, no exit 1 —
but `plan_lint --strict` fails a stop-less day, so give a move day its airport stop.
**Redirect sun's output to a file rather than piping it** (`… sun plan.geo.json
--write > sun.log 2>&1`, then read the file) — a pipe makes `$?` the *last*
command's exit and loses sun's non-zero signal. A transient TLS failure on one
day is expected, not breakage: re-run with `--only DATE` for the day it names.
Canonical `sun` format — the renderers parse it, so keep the shape:
`天亮 HH:MM · ☀ HH:MM / 🌇 HH:MM · TZ · sunrise-sunset.org`
e.g. `天亮 05:28 · ☀ 05:53 / 🌇 17:38 · JST · sunrise-sunset.org` —
`plan_lint --strict` accepts this shape only; never hand-write it.
(TZ may be a numeric offset like `-05` where the zone has no abbreviation — normal).
The dawn word follows the plan language: `sun --write` picks it from `--lang` >
`plan.lang` > `plan.meta.lang` > zh, so an `en` plan gets
`dawn HH:MM · ☀ HH:MM / 🌇 HH:MM · TZ · sunrise-sunset.org`
e.g. `dawn 05:28 · ☀ 05:53 / 🌇 17:38 · JST · sunrise-sunset.org`; the renderers
accept either spelling (zh output is unchanged).
**A space always follows a time**; never glue a bracket to it — `🌇 18:00(AEST`
is what the "golden hour ≈ …" margin line showed when a tester wrote
`18:00(AEST · …)`. Extra words go after a ` · ` separator. Trips that **cross a
daylight-saving switch** need a fetch on both sides of the switch (the script
does this per day; if you hand-fetch, take one date before and one after — the
Sydney 10-04 jump from 18:00 to 19:00 only shows up that way).
Manual fallback only when python itself cannot run (then `plan_lint` cannot run
either):
`curl -s "https://api.sunrise-sunset.org/json?lat={lat}&lng={lon}&date={YYYY-MM-DD}&formatted=0&tzid={Area/City}"`
— once per city (plus once per side of any DST switch), not per day; apply the
same sanity rules by eye (data-sources.md lists them); type the result in the
canonical shape above and stamp the day `note` "sun hand-fetched"; credit
sunrise-sunset.org in the footer either way.
The response carries `civil_twilight_begin/end` — that is 天亮/天黑 as a
traveller experiences it, ~25-30 min outside sunrise/sunset. Pre-dawn departures
and sunrise hikes schedule against **civil dawn**, not sunrise: a 06:00 trailhead
entry for an 06:22 sunrise is a dark-start (headlamp) only if civil dawn is 06:01.
Any day that starts before civil dawn gets the 天亮 time printed in its header and
a "prebooked car, no street-hailing in the dark" note. The mirror case at dusk: a
sunset viewpoint with no transit (a headland, a car-park lookout) is the hardest
place to get a ride — everyone leaves at once and the signal is often weak — so its
return ride is requested ~30 min early or pre-booked, the row names an alternate
pickup point, and nobody walks an unlit road or trail after dark.
Photography interest → schedule one viewpoint or walk at golden hour. Always check
evening blocks against sunset: an unlit garden at 19:30 in November is a bug; a
night-view deck is a feature — know which one you scheduled.
8. **Degradation plan**: tag every block that is neither `[pinned]` nor `[opener]`
as `[skippable]` or `[swap → alternative]`, and give each day one line: "running
>1 h late → cut X". On-the-ground plans fail; a plan without a failure mode is the
failure. Swap targets get the same closure check as anchors — a fallback that is
shut on the day it backs up is worse than none.
9. **Resolution**: quarter-hour granularity ("14:00-15:30") for anything you
estimated — false precision like "14:07" destroys trust in the honest numbers.
Verified timetabled departures and arrivals are the exception and keep their
published minutes ("09:15-11:31"), because that is exactly where the extra digits
carry real information.
10. **Weather shapes the day** — read the day's forecast or climate line
(data-sources.md §Weather) before ordering the blocks:
- **Heat** — max ≥ 32 °C, apparent max ≥ 35 °C or UV index ≥ 8 (UV exists only
in forecast mode; in normals mode a country note such as "UV extreme" is the
trigger): no unshaded
open-air anchor (ruins, markets, hikes) between 11:00 and 16:00 — they take the
opener slot or start after 16:30, and the early afternoon holds an indoor or
air-conditioned anchor or a long lunch; open-air dwell ×1.3 (the "hilly and
unshaded" surcharge above folds into it); the continuous on-feet cap drops from
3.5 h to 3 h; the day note carries "water · shade · hat".
- **Rain** — precipitation probability ≥ 60 %, or a rainy-season base (≥ 7 rain
days in 10 on the climate line): the `rain_alt` becomes the main line and the
outdoor version becomes the `[swap → …]`; every boat, cable-car and balloon
anchor carries "weather cancellation → backup date" in its note.
- **Wind** — gusts ≥ 50 km/h: balloons, boats, open decks and high towers get the
cancellation-risk note and a named fallback.
- **Cold** — min ≤ 5 °C, or min ≤ 0 °C at a dawn anchor (`temperature_2m_min`;
no recipe fetches an apparent minimum): shorter outdoor
blocks, lunch indoors, the dawn temperature printed in the day header note.
- **Visibility** — an anchor whose whole point is the view (a bridge or skyline in
a fog-prone coastal city, a crater rim, a summit): a clear point forecast is not
evidence, because fog and low cloud sit under it. The row links the official
live webcam or visibility page and names a go / no-go time (when the traveller
leaves the lodging) with its `[swap → …]`.
- **Smoke** — in a wildfire season (country-quick-notes.md §Hazard seasons), read
the air-quality index (US AQI ≥ 151 "Unhealthy"; ≥ 101 with kids or a
respiratory condition; elsewhere the local index's equivalent) and the park's
cameras and conditions page the day before — or by the free-cancellation cutoff
of a booking that hangs on the day, if that comes first. A hit swaps the outdoor
anchors for the `rain_alt`. Trip-level gates stay in phase-1-brief.md §Hazard
line.
The T-7 re-check (output-template.md §Pre-departure re-check ladder) re-applies
the forecast thresholds above; visibility and smoke are read the day before and
on the day. A gate days ahead (a forecast horizon, a hazard season) stays a
checklist row; a decision on the day or the evening before — the visibility go /
no-go, the smoke read — is a **decision point** in the day's `decision_points`
(shape: output-template.md §city-block), so `route_tools.py ics` never emits the
same gate twice.
## Day types that need a different structure
- **Arrival day** — landing before 15:00 = half a sightseeing day, later = **zero
sightseeing days for the count** — but the evening still gets one free, walkable,
unticketed block near the hotel (same rule in SKILL.md Phase 2 §3). Time the rest
of the arrival day from the **hotel door**, not touchdown (the 15:00 day-count
threshold stays on landing): landing + immigration (a late
international arrival at a big hub can take hours — use the airport's published
wait times or plan the worst case) + bags + transfer. The first morning after a
long-haul arrival, or after a hotel-door time past ~midnight, starts no earlier than
10:00. Plan the first evening within walking distance of the hotel; nothing
ticketed; dinner after a late arrival within ~500 m of the hotel, or inside it.
**Landing straight into a pinned activity** (a tour, a ride, a boat): write the chain
link by link — deplane 10-20 min + checked bags 20-40 min (the top end off a
wide-body) + the walk to the taxi / ride-hail zone (find its level; some airports put
ride-hail on departures) + the ride + the operator's check-in lead (rule 1). It fits
with 30 min to spare, or the activity or the flight moves.
- **Moving day (base change)** — luggage-encumbered from checkout to next check-in
(~15:00), so this is a day-structure problem, not a buffer problem. The day's first
row is timed from the booked option's `checkout` (`hotels[].options[]`, read off the
property's own page — SKILL.md Phase 5), else from the earliest checkout among the
options, marked est; a late check-out is asked the evening before and recorded on
that row. Solve bags before scheduling anything: (a) store at the departing hotel
(its `bag_hold` says until when) and loop back, (b) coin lockers or staffed storage
at the rail hub actually used — large lockers at major stations sell out by mid-morning, so name a fallback, (c) in Japan,
takkyubin hotel→hotel forwarding is usually **next-day**: arrange it the evening
before and keep an overnight kit. Never schedule an anchor with bags in tow — rule 2
(opener at opening time) is suspended on a checkout day unless bags are already
stored. **An intercity moving day carries ≤2 anchors, and only when the bags are
solved before the first anchor (checked / stored / hotel-held); otherwise 1** (same
sentence in references/phase-6-assemble.md §Adversarial self-check). Two is for the day whose train leaves at
16:00 and whose bags went into a locker at 09:00 (Göreme open-air museum + Dark
Church, then Uçhisar, then the night bus); a day that drags a suitcase to the first
sight is a one-anchor day whatever the timetable says. Third case,
**transfer-day-as-itinerary** — the bags ride locked in a private vehicle door to
door (a hired car with driver, a private tour van), never dragged and never stored:
then the anchor cap is the pace cap, not 2, and the surviving constraint is the
energy curve (rule 6). If the sunrise anchor is at
the *first* stop, set `sun_stop` (rule 7).
- **Departure day** — zero sightseeing unless the flight leaves after 18:00. Work
backwards from the airport-arrival deadline (3 h international) through the real
city→airport transfer time (the first row timed from the checkout as on a moving
day), and put the luggage solution in writing again.
**Return-flight time unknown** (`flight_scan.py` lists outbound legs only, and the
deep link may not show it statically): pick a plausible departure `T` from the
carrier's published schedule (or the outbound's mirror), and write the day
backwards from it as a formula the reader can re-run: **T = takeoff → T−3 h at the
airport (international; 2 h domestic) → T−4 h 30 leave the hotel** — e.g.
`T ≈ 12:30 → 09:30 at IST → 08:00 out of the hotel`. The `legs` row carries
`"dep": "⚠️ ~12:30 — verify"` (not a bare time), the timeline blocks say "T−4h30"
next to the clock, and `unverified` gets one line: "return flight departure time
not confirmed — day 10 timeline is built on T ≈ 12:30". A confidently wrong 08:00
taxi is worse than an honest formula. Where the country refunds tax at departure,
that kiosk step sits before bag drop and its queue is added to the airport deadline
(country-quick-notes.md for the country).
- **Tour day (跟团日)** — the pickup is `[pinned]` with a 10-15 min margin counted
from the **start** of the pickup window (14:00-14:30 → in the lobby by 13:45-13:50);
the exact time and point the operator sends later replace the window. The
operator owns the clock: no `late_cut` authority inside the tour, so the day's
degradation plan covers only what the traveller controls (meals, the evening, the
next morning's buffer). The evening BEFORE a tour day carries its own pins: buy
provisions, pack per the tour's luggage rule (2-day tours are often
overnight-bag-only), sleep early enough for the pickup. A "free activity" block
inside a tour = the bus waits for latecomers — schedule an alarm, not an itinerary.
- **Early start at a small lodging** — motels, inns and B&Bs often staff the front
desk only by day. When the day leaves before the desk opens (a pickup, an airport
run, a sunrise drive), the evening before carries the pins: settle the bill, agree
the key drop, store the bags that stay behind. Read the desk hours when the lodging
is shortlisted.
- **Safe-version day (精简安全版)** — when the traveller says a place feels
overwhelming or unsafe, offer this first, before any rebuilt full day: one anchor
(the one they would most regret missing), door-to-door rides for every hop, back
at the lodging before dark, on foot only around the hotel block; add blocks only
when they ask. Empty is not unsafe — business districts empty out on Sunday
afternoons and after office hours; say so in the day note, or give the ride, so a
quiet street does not cancel the outing.
- **Dateline day** — a westbound trans-Pacific return to East Asia consumes **two
calendar days** (depart HNL Oct 5 → land Oct 7); eastbound you land the same day
you left. Lock the return leg FIRST and allocate the final city's days from it —
planned the other way round, the dateline silently deletes a day from the last
stop (learned live: it cost Hawaii its circle-island day).
- **Weather-window day** — trading a booked flight for a clear day elsewhere. Read
every same-day flight the fare can change to, with its real change price, in the
carrier's app — fare buckets move within minutes, so a price seen is not a quote:
re-read it at the moment of changing. Nothing moves without two kinds of evidence:
the forecast (hourly, plus the forecaster's written discussion) and the present
(satellite, METAR, an official camera) — recipes in data-sources.md §Weather step
5. Decide the evening before and re-read the present at breakfast; both are the
day's `decision_points` (rule 10), each with its own `what`. The leaving day's clock re-derives from the
booked option's `checkout` (Moving day). The change goes to the traveller as a
priced option they choose (SKILL.md Interaction contract).
## Traps that break otherwise-correct timelines
- **Worship-hour closures — everywhere, big cities included.** Mosques close to
non-worshippers during the five daily prayers (~30-45 min, times shift daily) and
for most of Friday midday — pull the date's prayer times rather than assuming.
Churches close to tourists during Mass, Sunday mornings especially, and across Italy
and Spain they commonly shut 12:00-15:00 — **in central Rome, Milan or Florence the
shops stay open through the afternoon but the churches still close**, which is the
half of the story most plans miss. Every worship-site block carries a one-line dress
code (shoulders and knees covered; headscarf where required) — being turned away at
the door is a preventable failure.
- **Commercial siesta — Spain nationwide, Italy/Greece outside the big-city centres.**
Independent shops and family restaurants close roughly 13:30-17:00. The
mid-afternoon block there must be something that stays open (major museum, garden,
café terrace, long lunch), with shopping and the paseo moved to 17:30+. Assume
midday closure for any small venue in these areas unless its hours were verified.
- **Weekend food closures**: city markets are a classic Sunday trap (Rome's Testaccio
and Campo de' Fiori both shut). If food is a stated interest, verify the market's
own day-of-week before building a meal around it.
- **Closures the holiday API can't see**: local festivals (streets closed, hotels 3×),
seasonal operating windows (cable cars, mountain huts, gardens, boats), per-venue
annual maintenance, and Ramadan in Muslim-majority destinations. SKILL.md Phase 1
budgets a search for these; the Phase 6 closure scan re-checks them.
- **A boarding pass with no seat number is not a seat.** "See agent" or a seat
assigned at the gate means no seat is held yet: check in the moment check-in opens,
be at the gate desk as soon as it is staffed, and on a must-catch leg
(phase-3-legs.md §Record for every leg) treat it as a denied-boarding risk.
## Timeline verification (runs inside the Phase 6 self-check)
- Chain arithmetic: every block starts ≥ previous block end + hop time + buffer
- Anchors: arrival ≥ opening; arrival ≤ last entry; planned exit ≤ closing
- Pinned tickets: arrival margin matches the tier in rule 1 (15 vs 30-45 min, or the
operator's published check-in lead); a tour pickup's margin counts from the start
of its window (§Day types)
- No more than ~3.5 h of continuous on-feet anchor time without a sit-down block
- Walking total ≤ 8 km, counted honestly: `check`'s **on-foot** line ×1.3 (its ridden
line is not walking — mark those hops `"mode": "transit"` so it can tell), **plus**
in-venue walking and any stroll segments. A stroll is invisible to `check` unless
you model it as a stop at its midpoint — so model it, and show the arithmetic in
`walking_km: {total, how}` rather than asserting a bare number.
- Late evenings: last planned hop vs the line's last departure (verify, don't assume).
If lodging isn't in scope, say "day starts at {first block} / ends at {last block};
add your hotel hops" and mark this check N/A rather than passed. A sunset viewpoint
without transit has its early-requested or pre-booked ride and an alternate pickup
point (rule 7).
- Meal blocks exist and actually sit near the clusters they claim
- Moving days: first row timed from the booked `checkout` (or est); bags solved in
writing, no anchor before storage; ≤2 anchors only if the bags are solved before the
first one, otherwise ≤1 (§Day types)
- Departure day with an unconfirmed return time: the T-formula is visible in the
timeline, `legs.dep` carries ⚠️, `unverified` names it (§Day types)
- Worship/siesta/free-day traps checked for every affected block
- Weather rule 10 applied: a heat day has no unshaded open-air anchor 11:00-16:00, a
wet day's main line is the indoor version, wind-sensitive anchors carry the
cancellation note and a fallback, view anchors link their webcam with a go / no-go
time in the day's `decision_points`
## Scheduled-day format
Hop rows use the canonical format from navigation.md, and every hop is marked
`(verified)` or `(est.)` so the reader knows which durations were actually checked.
**The arithmetic below is meant to be copied**: every block starts at the previous
block's end plus the hop plus its buffer, and this day passes the whole verification
list above. If you catch yourself writing a 0-minute gap between two places 1 km
apart, that is the bug this example exists to prevent.
```
2026-10-05 (周一) — 京都·东山 天亮 05:28 · ☀ 05:53 / 🌇 17:38 · JST · sunrise-sunset.org
08:00-09:15 清水寺 ¥500 — 06:00 开门,首小时人最少 [opener]
09:15-09:45 步行 清水坂→三年坂→高台寺 0.7 km ~10分(+街景停留) (est.)
09:45-10:45 高台寺 ¥600 — 09:00-17:00,最晚入场 17:00
10:45-11:00 步行 高台寺→八坂神社 0.4 km ~5分 (est.)
11:00-11:30 八坂神社 免费 — 境内 24h [skippable]
11:30-11:45 步行 八坂神社→祇园四条 0.5 km ~8分 (est.)
11:45-13:00 午餐 · 祇园/花见小路一带 [swap → 锦市场]
13:00-13:45 巴士 203(往銀閣寺道) ~25分 ¥230 · 祇園→銀閣寺道 · 下车步行8分 (est.)
13:45-15:15 银阁寺 ¥1,000 — 08:30-17:00(3-11月)
15:15-16:00 哲学之道 散步 ~1.5 km(按站点建模,否则不计入步行总量) [skippable]
交通算账:巴士 2 段 × ¥230 = ¥460 < 地铁巴士一日券 ¥1,100 → 按次刷卡
步行合计:check 1.6 km(含祇园午餐点)×1.3 + 哲学之道 1.5 + 馆内 ≈ 5.0-5.5 km < 8 km ✓
迟到 >1h → 先砍八坂神社,再砍哲学之道;银阁寺须 15:30 前入场
雨备:三十三间堂(室内长堂,周一开放 — 已核实)。⚠️ 勿用京都国立博物馆:周一闭馆
时刻/票价为 2026-08-01 查得的季节性数据,出行前两周复查
```
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |