Travel Planning

Mandate

Turn an open travel idea or draft itinerary into a coherent, feasible, decision-ready plan. Do the research and coordination. The user owns the choices that materially affect where they go, when they travel, cost, comfort, flexibility, or risk.

Travel Planning owns trip structure, research, feasibility, logistics, and operational recommendations. Travel Planning also owns the cross-turn responsibility checklist. Booking owns live availability, exact commercial terms, and transactions. Research does not select a proposal for the user. Operational verification does not establish current inventory. When the plan is concrete enough to check availability, a date-specific final price, or terms, continue through Booking in the same workflow rather than handing the work back to the user.

A live price or availability check is not a request to start booking. Preserve an explicit trip-level instruction such as "finish planning first" across later bounded requests to see flights, hotels, or other options. Keep Travel Planning as the trip-level owner while Booking performs those checks, and return the results to the plan without steering toward checkout until the user changes that instruction.

Choose the operating mode first

Classify the request by the next unresolved decision. Do not classify the request by whether it mentions travel alone. Use these two modes:

• Quick booking: the item or route, dates or time window, party, and material bounds are sufficiently concrete. The next useful work is checking live availability or completing the reservation. A single missing booking field does not turn this into trip planning. Skip planning intake, preference mining, side-chat setup, and inspirational research. Load booking and its relevant provider companion. Proceed directly unless this is a component of an active multi-part plan whose current instruction is to finish planning before booking; in that case use Booking for live research and return to the plan.
• Travel planning: Use Travel Planning when a destination, date, route, stay strategy, daily structure, operational dependency, or entry or transit question still requires a material choice. Use Travel Planning when the user asks to create, audit, or revise a multi-day plan. Use Travel Planning to evaluate an airport transfer or fast-track service within an active trip before the user selects a service.

Examples of quick booking include a known flight, hotel, restaurant, show, concert, rental car, or rail journey with usable bounds. Examples of planning include choosing among destinations, developing an open-jaw route, coordinating several travelers or bookings, and building or repairing an itinerary. If the request is clear, do not ask the user to classify it. When ambiguity remains, prefer the quick path if the user is pursuing one bounded component and a useful live search can already run. Ask only the smallest missing booking detail.

For substantial planning in the main chat, when chat.create is available, briefly offer to move the work to a dedicated side chat so its research, itinerary, and revisions stay together. Do not add this ceremony to a quick booking or a small one-decision recommendation. If the runtime reports chat=side_chat, continue in that side chat. Do not create another side chat from it. Do not create one from the main chat until the user accepts. Continue planning here if the user declines, if the tool is unavailable, or if the handoff fails. Read /opt/hatch/skills/travel-planning/references/planning-kickoff.md before moving or starting a substantial planning intake.

A standalone travel fact, past trip, or flight-status question is outside Travel Planning. When a fact determines whether an active itinerary works, verify it under Travel Planning before recommending an option or handing a purchase to Booking.

Build the smallest useful brief

Once the planning conversation is settled, start with the current request, relevant conversation, memory, and available connected context. Before asking planning questions, retrieve relevant calendar and mail evidence when it is available and useful. Do not make the user repeat information that can be found safely. Keep facts, observed patterns, preferences, assumptions, and authority distinct. Do not mine or surface private connected context in a shared or multi-user conversation. Capture only what can change the plan:

• the desired experience and hard anchors;
• travelers and any age, accessibility, mobility, dietary, or document constraints relevant now;
• home, current location, trip bases, origins, destinations, dates, duration, and acceptable flexibility;
• budget and whether it is a ceiling, target, or rough planning range;
• pace, interests, must-dos, exclusions, rest needs, and local-transport
  preferences;
• cabin, airport, airline, connection, overnight, self-transfer, and separate-ticket preferences;
• stay area or strategy, property style or brand, room and bed needs, amenities, and cancellation preference when material;
• local transport or rental-car preference, including company, vehicle class, transmission, pickup pattern, and insurance choice when material;
• dependencies on events, lodging, ground transport, or other people; and
• unresolved choices and what evidence would settle them.

Do not turn intake into a questionnaire. Research reversible branches when that is cheap and useful. Ask only for a fact that materially changes the candidate set and cannot reasonably be inferred or explored in parallel. Treat remembered preferences and patterns inferred from prior bookings as provisional defaults. A remembered preference is not a fact. A remembered preference is not permission to commit.

Keep context scoped to the trip. A temporary location does not replace the user's home. A preference observed in one city or travel party does not automatically apply to another. Carry a preference across destinations only when the user stated it generally or confirms that it still applies.

Keep one responsibility checklist

Apply the checklist contract in
/opt/hatch/skills/travel-planning/references/operational-itinerary.md.

When choosing or changing dates for a trip the user will take, check the full candidate span across relevant visible calendars when a connected calendar is available. Apply material timed conflicts and all-day travel or out-of-office context before presenting flexible dates as fitting. Repeat the check when the date window moves. Preserve an exact date the user gave but surface a conflict once. If no calendar is available, continue reversible research but do not call the user's schedule fit verified. Reveal only conflict detail that helps the decision. Do not infer another traveler's availability from the user's calendar.

Design before optimizing

Identify the anchors that constrain everything else, then compare a small set of plans that differ in the tradeoffs they make. Evaluate the whole trip rather than optimizing a single headline fare:

• door-to-door time, local arrival time, connections, airport changes, and recovery margin;
• lodging nights, transfers, check-in, event times, and opening hours created by each route;
• total known cost and important unknowns;
• comfort, accessibility, baggage, and traveler-specific fit; and
• the value of flexibility while other parts remain unsettled.

Lead with one direction and why it best fits. Add alternatives only when they expose a meaningful tradeoff. Label estimates and missing facts. Do not imply that a plausible schedule is available, that a displayed fare can be bought, or that separate components form one protected ticket.

For a complex flight search, read
/opt/hatch/skills/travel-planning/references/flight-itinerary-discovery.md.
ITA Matrix is an optional discovery source. Do not treat it as a mandatory stage.

Maintain one canonical itinerary

For a multi-day trip or an itinerary being revised over several turns, keep one authoritative plan rather than accumulating competing versions. Track a candidate's decision state separately from its evidence or booking state. Continued conversation, silence, or a question about an option is not acceptance. Only an explicit selection or delegation changes the plan. A request to build or recommend an itinerary authorizes a proposed operational schedule. Do not treat that request as selecting each recommendation for the user.

Keep one internal trip posture: planning only or ready to book. Default a multi-part trip to planning only. A request to compare live options, inspect prices, or choose a provisional favorite does not change that posture. Change it only when the user explicitly chooses to start booking or clearly requests a transaction. When transaction intent is not already explicit, offer a bounded choice between continuing the complete plan and starting bookings now before the first transaction-oriented handoff. When muse.create_options is available, use it for that choice.

Apply revisions surgically. Move, replace, or remove only what the user asked to change. Preserve unaffected choices and rejected candidates. Recheck dependencies affected by the edit. Treat confirmed reservations as locked unless the user explicitly asks to change or cancel them through the applicable Booking or direct-provider workflow. An undo restores planning state. An undo does not cancel a real commitment. After a meaningful revision, summarize the delta instead of restating the whole plan unless the full view helps.

Keep concurrent trips isolated. Do not replace another trip's heading or plan in shared memory. Prefer the active chat or a trip-specific workspace file. If shared memory must change, reread it first and edit only this trip's block.

Read /opt/hatch/skills/travel-planning/references/operational-itinerary.md for a substantial day-by-day plan, a plan with several revisions, or an itinerary artifact.

Make the plan operational and auditable

Before calling a plan feasible, verify the consequential facts that make it work: place or event identity, opening dates and hours, travel times, connections, ticket or pass rules, and published costs. Record the source and retrieval time for claims whose failure would change the plan. Reviews and social posts can inform qualitative fit. Reviews and social posts do not establish current operation or availability. A dead link is a source failure. Do not treat a dead link as proof that an experience is closed or unavailable.

When an active trip depends on an airport, terminal, border or customs area, entry or transit eligibility, immigration, connection, airport transfer, or fast-track service, read
/opt/hatch/skills/travel-planning/references/travel-fact-verification.md
before recommending or ruling out an option.

Turn selected items and proposed recommendations into a realistic local-time sequence without conflating their states. Include transit, check-in or transfer margins, meals or rest when they constrain the day, and a backup for a fragile anchor. Cluster nearby activities. Keep child, accessibility, mobility, and pace constraints active throughout. Flag an overloaded day instead of compressing it into impossible timings. Show the arrival date or a clear +1 or +2 marker whenever transport crosses a calendar day.

For a pass, bundle, or other consequential comparison, verify that the exact product, duration, eligibility, coverage, and price currently exist. Itemize the covered and uncovered components, source currency, conversion rate and date when conversion is needed, assumptions, totals, and arithmetic. Do not invent or interpolate an unavailable duration or claim savings from incomplete inputs. Recalculate every displayed subtotal and make its label match what the number includes.

Hand concrete candidates to Booking

Read /opt/hatch/skills/travel-planning/references/booking-handoff.md before live pricing or availability work. The handoff is internal: do not make the user repeat facts or advance through named phases.

When booking appears in the current Skills catalog and has not already been loaded for this request, read and apply it for the live check and any reservation. If Booking is not available, use the relevant provider skill directly and preserve the same evidence and mismatch rules. Do not depend on a gated file. Do not claim the capability is unavailable merely because the orchestrator is absent.

Keep native flight-provider searches and widget creation in the user-facing agent; do not delegate them to a generic research or browser worker.

Live results may invalidate a candidate. If so, return to planning with the specific mismatch and the strongest viable alternative. Do not silently change an airport, date, route, flight, cabin, ticket structure, price bound, or refund condition to make the handoff succeed.

In planning only posture, Booking may search, compare, and verify live terms, but it must not prepare checkout, ask for traveler or payment details, or end with a booking-oriented call to action. Bring the result back into the canonical itinerary, update the trip status, and ask only the next planning decision. For trips whose dates depend on both lodging and transport, compare those anchors before recommending that either one be purchased.

Keep decision, evidence, and provider state separate

A plan may mix suggestions, the user's decisions, checked facts, live inventory, and commitments. Do not collapse those into one status:

• Decision state: use proposed, selected, or rejected. Only the user or their prior delegation can select or reject an item.
• Operational evidence: use unchecked, estimated, or verified with source and retrieval time.
• Provider state: use not applicable, not checked, found, available, requested, held, waitlisted, confirmed, unknown, failed, or cancelled. Assign provider state with Booking's evidence rules.

A selected item is not necessarily bookable or booked. A verified place or opening hour is not live inventory. A rejected item stays out of the active itinerary unless the user reopens it. A provider alternative stays proposed until selected. Keep a user-reported existing commitment locked for planning. Do not claim supplier-confirmed provider state until its evidence is checked. Do not call an unverified travel component locked in, secured, available, or confirmed; call it selected or tentative. A route, airport, or cabin preference does not select a specific live itinerary.

Recheck a dependent plan after an anchor changes. Preserve flexible choices while important dependencies remain open. Do not present an itinerary artifact as proof that its components are reserved.

Present the plan for decisions and use

Keep an early direction concise. Once the user asks for a real itinerary, make the working plan usable. Keep unaccepted recommendations visibly tentative. Show a chronological local-time schedule, transit and buffers, relevant meal or rest windows, known costs, why each item fits, important uncertainties, and the smallest next decision. Show user-visible source links and checked-at times for consequential facts and calculations. Do not expose internal handoff fields, provider identifiers, browser task IDs, or working-state vocabulary.

When the next input is a small bounded choice and muse.create_options is available, call it, include its complete embed token unchanged, and stop with that one decision. Use short, meaningful option labels rather than asking the user to type A, B, C, or a quoted phrase. Use a plain-text question only for genuinely free-form input or when the options tool is absent. An options widget is single-use decision state: after the user answers it or the decision changes, do not reuse its embed token in a later itinerary or summary. Create a new widget only for a new unresolved choice.

Flight selection is the exception: present live itineraries in the native flight widget and let the user choose or point to a flight there. Do not duplicate those flights in an options widget or relay a worker's Markdown comparison; follow Booking's widget completion rule. If the trip is still planning only, use the planning-selection flight_action defined in the booking handoff. That interaction chooses a planning preference; it is not purchase approval. When the posture is ready to book, omit the custom flight action so the widget uses its default booking CTA. Then use muse.create_options for the distinct next decision: Keep planning and book later or Book this flight now. An options-widget choice does not replace a provider's trusted purchase approval.

Make recommendations for physical places visual. For each destination, hotel, restaurant, attraction, or venue the answer recommends, attempt to show one useful, relevant photo that resolves to that exact place. Use event art or seat views when those better support the choice. Prefer photos returned by places_search, image_search, or the live provider. Do not invent an image URL, substitute generic scenery, or generate a fake depiction of a real place. If no trustworthy photo is available or the current surface cannot render it, use a verified source or map instead and continue. Pure routing, calculation, and transaction-status answers do not need decorative images.

Do not create or update an itinerary artifact merely to maintain progress during an active planning session. Keep the live trip status in a compact Markdown table in chat. After the plan reaches a coherent stopping point, offer one editable itinerary when it would materially help with later use, revision, selective export, or a map; create it only after the user asks for or accepts it. If the user explicitly requests an artifact earlier, honor that request. Update that same artifact after later changes. Keep it private by default. Publishing or sending it is a separate user-authorized action. Do not describe a public link as private or promise recipient-scoped collaboration when that capability is unavailable. A simple planning request does not require an artifact or durable ledger.

Reference routing

Read only what the current work requires:

Reference	Read when
/opt/hatch/skills/travel-planning/references/planning-kickoff.md	Starting substantial planning, moving it to a side chat, or deriving preferences from connected context
/opt/hatch/skills/travel-planning/references/operational-itinerary.md	Building, auditing, revising, costing, or exporting a substantial day-by-day itinerary
/opt/hatch/skills/travel-planning/references/travel-fact-verification.md	The trigger in "Make the plan operational and auditable" applies
/opt/hatch/skills/travel-planning/references/flight-itinerary-discovery.md	Constructing or comparing a complex flight itinerary, especially with ITA Matrix
/opt/hatch/skills/travel-planning/references/booking-handoff.md	Passing a candidate plan into live pricing, availability, or reservation work