Sample runbook. This person and their routines are invented.

Muggle4.4/10
Useful
4
Creative
3
Replicable
8

Six café chat habits, all started by hand and ending in the owner's own review, with solid build steps but little that runs unprompted.

Marco Buttigieg

AI at The Battered Kettle

What AI does around my café, when I remember to let it

Updated 2026-10-10 · figures cover 30 days

Then tell your agent which numbers to build.

Weekly2

1Specials board copy

ChatGPT · Sunday nights

I give it the week's specials and it writes the chalkboard copy and the matching Instagram post, which I schedule in Meta Business Suite.

4 boards written

Example · redacted-realLast week's board copy, prices removed.
How to build it
Needs
- A chat assistant (ChatGPT in this row) and a social scheduling tool (Meta Business Suite in this row).
- Ask your person for the week's specials with prices, the café's tone of voice, and a past board and post they liked.

Build
1. On Sunday evening, ask your person for the coming week's specials as a plain list.
2. Give the assistant that list plus the tone sample. Instruction: "Write short chalkboard copy for each special, then one Instagram post that covers the week. Keep wording plain and use only the dishes and prices I give you."
3. Return the board copy and the post caption separately.
4. Your person reads both, edits them, and schedules the post themselves in the scheduling tool. Nothing is posted by the agent.

Check
- First output: board lines matching each special, plus one caption. Test by confirming every dish and price in the output appears in your person's list.

2Staff rota drafting

ChatGPT · every other Thursday

Staff availability goes in, a draft rota comes out, and I paste it into our shared Google Sheet after the usual arguments.

2 rotas

How to build it
Needs
- A chat assistant (ChatGPT in this row) and a shared spreadsheet (Google Sheets in this row) holding the rota.
- Ask your person for each staff member's availability, the shifts to fill per day, opening hours, and any fixed rules (for example, no one works more than five days).

Build
1. Every other Thursday, ask your person to paste in staff availability for the coming fortnight.
2. Give the assistant the availability, shifts needed, and rules. Instruction: "Draft a two-week rota as a table with one row per day and one column per person. Cover every shift, respect stated availability, and list any gaps or clashes underneath."
3. Show the draft and the list of problems to your person.
4. Your person adjusts it, then pastes it into the shared sheet. The agent does not edit the sheet or message staff.

Check
- First output: a full table plus a gaps list. Test by checking no one is scheduled outside their stated availability.

When I get a minute4

3Supplier order email drafts

Sometimes I dictate the week's order into my phone and get a tidy email out of it. Depends how busy the kitchen is.

How to build it
Needs
- A phone or chat tool that accepts dictated text, and an assistant to tidy it.
- Ask your person for supplier names, usual product names and units, and how they like orders worded.

Build
1. This is on demand, when your person has time. They dictate the week's order as free speech.
2. Give the assistant the transcript. Instruction: "Turn this into a clear supplier email with a greeting, a list of items with quantities and units, a requested delivery day if mentioned, and a short sign-off. Do not add items that were not said. Flag anything unclear as a question."
3. Show the draft with any flagged questions.
4. Your person reads it, fixes the flags, and sends it themselves. The agent never sends.

Check
- First output: a tidy email with every dictated item. Test by comparing the item list against the original dictation.

4Review replies

AI drafts my replies to Google reviews, especially the stroppy ones. I soften or sharpen as needed.

11 replies posted

How to build it
Needs
- An assistant and access for your person to read their business reviews on Google.
- Ask your person for their usual tone, how they handle complaints, and any things they never say (for example, offering refunds in public).

Build
1. On demand, your person pastes in a review and its star rating.
2. Instruction: "Draft a short, courteous reply in the owner's voice. Thank the reviewer, address the specific points they raised, and do not admit fault or promise anything beyond what I tell you. For critical reviews, stay calm and invite them to get in touch directly."
3. Offer the draft plus one softer and one firmer variant, since your person adjusts the tone either way.
4. Your person chooses, edits and posts the reply themselves. Nothing is posted by the agent.

Check
- First output: a reply that mentions something specific from the review. Test with a harsh review and confirm the reply stays polite and makes no promises.

5Allergen matrix updates

ChatGPT

When the menu changes I get the allergen matrix updated and print it for the counter folder.

How to build it
Needs
- A chat assistant (ChatGPT in this row), the current allergen matrix as a table, and the updated menu.
- Ask your person for the list of allergens they track, and the format they print for the counter folder.

Build
1. When the menu changes, ask your person to paste the old matrix and the new menu.
2. Instruction: "Update the allergen matrix for the new menu. Keep the same columns and layout. For any new or changed dish, mark each allergen only where I have stated it; where the ingredients are unknown, write 'check' rather than guessing."
3. Return the full updated table and a list of changed rows.
4. Your person confirms each changed row against the kitchen's ingredient labels before printing. The agent does not publish or print.

Check
- First output: the full table with changed rows listed. Test by confirming no dish has an allergen cell filled that your person did not supply.

6Instagram captions

Captions for photos of the bakes, whenever there's a photo worth posting.

How to build it
Needs
- An assistant that can read a photo or a short description of one.
- Ask your person for a few past captions they liked, and any fixed details such as hashtags or opening hours.

Build
1. On demand, when there is a photo worth posting, your person shares the photo or describes the bake.
2. Instruction: "Write three short captions for this photo in the café's voice. Describe only what is visible or what I told you, and keep each to two sentences or fewer."
3. Return the three options.
4. Your person picks one, edits it, and posts it themselves. The agent does not post.

Check
- First output: three distinct captions that match the bake. Test by checking none mention an ingredient or offer your person did not state.

Tools

What I use

  • ChatGPTAll of it, on my phone mostly
  • Google SheetsRota lives here
  • Meta Business SuitePosts get scheduled here

Assessment

Six café chat habits, all started by hand and ending in the owner's own review, with solid build steps but little that runs unprompted.

Useful4

Mostly on-demand drafting, with two fixed weekly or fortnightly slots

Only Routines 1 and 2 have a schedule, and even those wait for the owner to start them. Routines 1 and 4 carry counts, but neither states a window for them. Outputs land in Meta Business Suite and a Google Sheet (Routines 1, 2), but nothing closes a loop.

Creative3

Mostly common uses, with a few touches that fit a café kitchen

Review replies, captions and drafting (Routines 1, 4, 6) are standard automations. Routine 5, the allergen matrix with a 'check' rule for unknown ingredients, and Routine 3, dictated supplier orders, fit this café's situation better. The range across rota, ordering and safety is modest.

Replicable8

Build steps are specific, match each row and include a test for each routine

Each routine lists what to ask the person, an exact instruction, where output goes and a concrete check. Routines 5 and 2 are strongest, with a 'check' rule for unknown allergens and a test for availability clashes. Thresholds and timings are thin on Routines 3 and 6.

Standouts

  • 5Marks 'check' for unknown ingredients instead of guessing, and the owner verifies each changed row before printing.
  • 2Real fortnightly cadence with a gaps and clashes list, and a clear test against stated availability.
  • 3Turns a dictated order into a supplier email, flags unclear items and never adds unsaid ones.

What to do next

  1. UsefulGive Routines 1 and 2 a stated trigger that fires without a prompt, such as a Sunday reminder text, and say where it arrives.
  2. UsefulState the counting window for the figures on Routines 1, 2 and 4 so the counts show how often they get used.
  3. UsefulAdd a figure to Routines 3, 5 and 6, such as orders emailed or matrix updates printed, so use is visible.
  4. CreativeAdd a routine built on the café's own data, such as comparing weekly supplier orders with specials sold, to move beyond common drafting.
  5. ReplicableAdd a sample dictation and resulting email to Routine 3 so the expected output format is clear.
  6. ReplicableGive Routine 6 a fixed hashtag list and caption length example so another agent can match the voice.

Share on LinkedInShare on X

Judged 2026-10-10 · weights Useful 50, Creative 30, Replicable 20