Rosa's Runbook

Needs
- A reservations system that can export tomorrow's bookings (party size, dietary notes, guest tags).
- A Zapier flow, and a source of past covers per weekday.
- Ask your person for the dish list, portions per cover for each dish, and who reads the sheet.

Build
1. Schedule the flow for 9pm daily.
2. Pull tomorrow's bookings: covers, party sizes, dietary notes, regulars and occasions.
3. Pull seated covers for the last four same weekdays and average them.
4. Instruct the model: "Scale each dish's portions to the average of the last four same weekdays, adjusted for tomorrow's bookings. List dietary notes and occasions by table. Show your arithmetic."
5. Write the result to a one-page prep sheet and print it for 7am.

Check
- First sheet should show booked covers beside the four-weekday average, with portion counts per dish.
- Test: compare a week of portion guesses with what was actually used; flag any outside 10%.

Needs
- Access to new reviews from each review platform, via email alerts or an export.
- ChatGPT with a sample of your person's own replies loaded as a voice guide.
- Ask your person for 5 to 10 past replies they like and any topics never to promise.

Build
1. Schedule a daily run at 8am that collects reviews posted since the last run.
2. For each review, instruct the model: "Reply in the owner's voice using the samples. Warm and brief for positive reviews. For negative ones, name the specific complaint, own what is true, state what changed, and do not grovel or offer compensation unless told to."
3. Put each review and its draft in an approval queue.
4. Your person taps approve, edits or skips. Nothing posts without that approval, and do not run any drafting late at night.

Check
- First run should give one draft per new review, each mentioning a detail from that review.
- Test: feed in a 1-star review and confirm the draft is specific and calm, not defensive.

Needs
- A kitchen tablet form builder and Zapier (or similar) to collect entries into a sheet.
- Ask your person for the reason list and the monthly meeting date.

Build
1. Build a short form with fields: item, time (auto-filled), reason (ran out, quality pulled, prep short), and optional quantities sold and prepped.
2. Put the form on the kitchen tablet and ask staff to fill it in at close of service.
3. Send each entry through Zapier into a single log sheet, one row per item.
4. At month end, add a tab counting entries by reason and by item, and highlighting any item 86'd on the same weekday three times running.
5. Keep the log read-only for everyone except the kitchen lead.

Check
- The first night should add rows with time and reason and no blank fields.
- Test: log a dummy item with prepped and sold numbers and confirm the repeat-item highlight triggers on the third entry.

Needs
- Supplier invoices arriving as PDFs, forwarded to one mailbox or folder.
- Zapier plus ChatGPT able to read PDFs.
- Ask your person for the flag threshold (default 5%), supplier contact details, and a tone sample for queries.

Build
1. Schedule a run for Monday mornings over the previous week's invoices.
2. Extract line items: supplier, product, unit, unit price.
3. Compare each line with its average over the last eight weeks. Flag any rise above 5%.
4. Instruct the model: "For each flag, state the old and new price and the percentage. Draft a short, polite query asking whether the rise is seasonal or an error."
5. Put the flags and drafts in an approval queue. Your person sends, edits or drops each one. Nothing goes to a supplier unapproved.

Check
- First report should list flagged lines with old price, new price and draft.
- Test: alter one price in a sample invoice by 10% and confirm it is flagged.

Needs
- A reservations system that can export bookings with party size, day, booking source and outcome (arrived, no-show, late cancel).
- A Zapier flow.
- Ask your person for their current deposit rule and what counts as a late cancel.

Build
1. Schedule the flow for Tuesdays, covering the last 30 days.
2. Export all bookings with outcomes.
3. Group no-shows and late cancels by weekday, party size band (1 to 2, 3 to 5, 6+) and booking source.
4. Instruct the model: "Report the no-show rate for each group with counts, for example 5 of 9. Say which groups carry the pattern and which are clean. Suggest a deposit rule that applies only to groups with the pattern."
5. Send the report to your person. Any change to deposit policy is their decision.

Check
- First report should show rates as counts beside percentages per group.
- Test: confirm every no-show lands in exactly one group per grouping.

Needs
- A staff group chat that can be exported or read by a digest tool, and a current rota.
- Zapier or similar for the digest.
- Ask your person for rota rules (minimum cover per shift, hour limits) and the casual cover list.

Build
1. Schedule a Wednesday noon run for the following week.
2. Collect gaps in next week's rota and swap or time-off requests from the chat since the last digest.
3. Instruct the model: "For each gap or request, propose one resolution that keeps minimum cover and respects hour limits, say why, and name an alternative. Use initials only."
4. Compile one digest, one line per item, each ending with 'Approve?'.
5. Your person approves the set in one go. Only then update the rota and tell staff.

Check
- The first digest should cover every open request and gap, each with a proposal.
- Test: add an impossible swap and confirm it is flagged rather than approved.

Needs
- A shared folder where the kitchen drops photos during the week.
- ChatGPT with caption examples in your person's voice.
- Ask your person for the tone, hashtags to avoid, and anything off limits (guests, faces).

Build
1. Schedule a Thursday afternoon run over the week's new photos.
2. Instruct the model: "Group the photos into three posts: the special, a supplier story, and a team moment. Write one caption each in the owner's voice, short and concrete, with no hype. If fewer than three themes exist, make fewer drafts."
3. Save the drafts with their photos in a review folder.
4. Your person picks, fixes captions and posts manually. Nothing posts automatically, and a week with no posts is fine.

Check
- First run should give up to three drafts, each tied to named photos.
- Test: run it with one photo and confirm it makes one draft, not three.

Needs
- A POS export of sales by dish, a per-dish cost sheet from the bookkeeper, and the 86 log.
- ChatGPT able to read spreadsheets.
- Ask your person for what counts as high-selling and high-margin (default: top and bottom third).

Build
1. Schedule the first Monday of each month, covering the previous month.
2. Join sales, margin per dish and 86 counts by dish.
3. Place each dish plainly: sells a lot or not, makes money or not.
4. Instruct the model: "For each dish write one line on where it sits. For any dish proposed for removal, state any reason to keep it from the data. Propose, do not decide."
5. Produce one pack for the kitchen meeting. The kitchen decides every change.

Check
- First pack should list every dish with sales rank, margin rank and a placement.
- Test: check that a dish 86'd often is noted as constrained, not as a weak seller.

Needs
- A mailbox or phone folder that receives receipts, and the bookkeeper's folder structure.
- A Zapier flow with a document reader for photos.
- Ask your person for the folder layout, the bookkeeper's rules for flagging, and the month-end day.

Build
1. Schedule the flow for the last Sunday of the month.
2. Collect emailed receipts and photographed paper ones from the month.
3. Read supplier, date, total and tax number from each, and file it in the bookkeeper's folder by month and category.
4. Flag a receipt with a missing date, total or tax number, or no category match, and list it in a short note.
5. Share the flag list with your person. Do not delete or rename any original.

Check
- First run should leave every receipt filed or on the flag list, with counts matching the inbox.
- Test: add a receipt with no tax number and confirm it is flagged.

Needs
- The kitchen tablet waste entries in a sheet, with item, weight and cause.
- ChatGPT able to read the sheet.
- Ask your person for the cause categories and where prep rules are kept.

Build
1. Schedule a run on the last day of each month.
2. Group entries by cause: over-prep, spoilage, mistakes.
3. Total each group in kilos and name the three biggest line items.
4. Instruct the model: "For each of the three, say which weekday or dish is involved and propose one rule for the prep sheet that would have prevented it, with a number."
5. Send the summary to your person. Any rule change to the prep sheet needs their approval first.

Check
- First summary should show kilos per cause, three named items and three proposed rules.
- Test: confirm the category totals add up to the sheet's total weight.

Needs
- ChatGPT with dish costings, drinks assumptions and deposit terms loaded.
- Ask your person for the three price points, minimum spend rules, deposit amount and standard room hours.

Build
1. When an enquiry comes in, paste its date, guest count and any dietary needs into the chat.
2. Instruct the model: "Build three menu options at the stated price points from the real dish costs. State drinks assumptions, add the deposit terms, and leave a blank sentence for the owner about the date."
3. Show the working: cost per head, margin and assumptions.
4. Your person checks numbers, fills in the date sentence and sends the quote. Nothing is sent to the enquirer without that approval.

Check
- First quote should have three options, each with a per-head cost and a total.
- Test: change a dish cost in the sheet and confirm the quote updates.

Needs
- ChatGPT with your person's phrases and an example of an ad they like.
- Ask your person for the role's real hours, pay range, benefits, and what a hard Saturday is like.

Build
1. Trigger it when a team member hands in notice, naming the exact role.
2. Give the model the hours, pay and duties.
3. Instruct the model: "Write the ad in the owner's voice. State the hours and pay plainly. Describe the hardest parts of a busy shift honestly in specifics. No hype, no vague perks."
4. Return one draft with a short list of claims to verify.
5. Your person edits and posts it. Nothing is published without their approval.

Check
- First draft should include pay, hours and one concrete hard part.
- Test: remove the pay range from the input and confirm the model asks for it rather than inventing one.

Needs
- ChatGPT on a phone.
- Ask your person for who is eating, foods the kids refuse, allergies, kitchen limits, and what leftovers came home.

Build
1. On request, tell the model who is eating on Sunday (default six), what the children will eat, and any leftovers from the restaurant.
2. Instruct the model: "Plan one family dinner, not a restaurant meal. Use the leftovers, nothing needing plating, and suggest a main, a side and a pudding. Keep the shopping list to what fits one basket."
3. Ask it for the list grouped by shop section and a rough running order.
4. Your person edits the plan freely before shopping.

Check
- First plan should have a main, sides, the leftovers used and a basket-sized list.
- Test: add a new allergy and confirm no dish or list item breaks it.