Prompt patterns that work

The AI can drive your dispatch board. Whether it does it well comes down to how you ask. This page collects patterns that consistently work, with explanations of why.

Always say "look it up first" when it matters

The single highest-leverage instruction. Without it, the assistant will sometimes generate plausible-sounding appointment details that don't exist.

Use appointments-list before saying anything specific about today.

You usually only need to say this once per session. Most assistants remember within the same chat.

Scope by day, not by week

Asking "plan my whole week" works but produces fuzzy results. The model juggles too many constraints at once and loses track of which appointments it already considered.

Better:

Plan tomorrow.

Then, once that lands:

Now Wednesday.

Same context, smaller bites, better decisions.

Be specific about who, where, and how long

The model isn't a mind reader. These three details unlock real planning:

Book Tom for the Verlinden job tomorrow at 10 for 90 minutes.

If you leave any out, the assistant should ask. If it doesn't, push back: "which Tom?", "what duration?". Treating it like a junior dispatcher is the right calibration.

Use the constraint, don't hide it

If the day has a rule, say so:

Avoid scheduling Eva before 09:30 tomorrow. She has a school run.

The assistant will use availability-suggest with the constraint baked in. Don't make it infer constraints from a previous turn five messages back. Constraints work best stated up front.

Move work in bulk

Where chat wins decisively over drag-and-drop:

Storm forecast Friday. Move every outdoor job to next week, keep indoor work where it is.

The assistant lists Friday's appointments, classifies outdoor vs indoor (using the kind/notes fields), proposes a reshuffle, and asks before committing. One sentence does what would take ten drags on the board.

Promote from the queue atomically

When a customer calls and you want to book a queued request:

Pull the Vermeulen request from the queue and book it Thursday morning.

The assistant calls appointments-create with from_queue_item_id. The queue item is deleted in the same transaction, so you never end up with both a queued and a booked version.

Ask for the report you want, not the data

How busy is Tom next week?

Returns a summary: hours booked, appointment count, slots remaining. The assistant does the listing and the math, you read one sentence.

vs

List all of Tom's appointments next week.

Which dumps a table. Both work; the first is faster for a human, the second is better if you'll process the data yourself.

Confirm before destructive moves

The assistant will usually confirm before mutating, but you can also force it:

Don't change anything yet. Tell me first what you'd do.

Then a follow-up: go ahead, or specific tweaks. Useful for high-stakes reshuffles.

Patterns that don't work

• "Plan optimally." Optimal for what? The assistant will pick a definition and surprise you. Be specific: "Plan to minimize driving", "Plan to keep Tom under 7 hours", "Plan to cluster by postcode".
• Mid-conversation tool switching. If you ask the assistant to write you an email mid-dispatch, it may evict the MCP tools from context. Start a new chat for the email.
• Vague references three turns back. "Like we did Tuesday" rarely works. Say what you mean explicitly.
• Implicit dates. "Next Friday" is ambiguous near a week boundary. Use the actual date when in doubt.

When to drop the chat

Chat wins for bulk operations, conditional logic, and queries. The board still wins for:

• Spatial edits (drag this card to that lane)
• Pixel-precise timing (move start to 14:15, not "earlier")
• Browsing without a goal in mind

Use whichever fits the task. They share the same data, so nothing is wasted.