Hangworkshop
step 3 · verde cantina · lapsed

Set the rules, then generate

Tune the config, attach the files, and run it once.

Attach verde-cantina-purchase-history.csv and verde-cantina-loyalty-members.csv from the last step, then paste the prompt.

Fill in the controls below — we’ll write the prompt for you. No markdown to edit; just set what you want and copy.

What the AI optimizes everything around — both the message and the send time serve this.

Generation — a unique message for every guest, written from scratch (pure 1:1).

words

Hard cap on the phrase.

A specific job for the phrase beyond the goal — e.g. a warm greeting, make them feel valued, reintroduce a favorite. Leave blank to let the goal lead.

Jot notes however rough — tone, vibe, words you love or ban, a line you'd write yourself.

One URL per line — your site, Instagram, a past email. The AI reads them to match your voice.

open in··
What you’ll get back
  • The upload file (main output). One CSV — every guest’s message, predicted send date & time, and an AI-written message rationale and send-time rationale citing that guest’s actual orders. This is the file you ship.
  • A template table (only in matching mode). If you chose matching, you also get the handful of templates to eyeball first. Want changes? Edit them and send them back to the AI to regenerate the upload file.
Why this prompt is built this way2026 practice
  1. 01

    XML-tagged sections

    role · campaign · inputs · config · constraints · process · output — the model can tell a command from data, so it follows instructions far more reliably than with prose.

  2. 02

    Raw data, not aggregates

    we attach the transactions, not a “customer profile.” Pre-summarizing is lossy compression — it throws away the exact behavior we want to personalize on.

  3. 03

    Constraints in their own block

    guardrails are isolated and explicit (banned words, never-mention, send window) — the highest-impact section for consistent, safe output.

  4. 04

    Three diverse examples, not one

    the output shows three intentionally different rows. A single example makes the model overfit and copy its style; a varied set pins the format while leaving room to be original per guest.

  5. 05

    Reasoning as a field, not a preamble

    we ask for a message- and timing-rationale per guest instead of “think step by step.” Modern models reason better when the structure forces it — and you get an audit trail.

Presented at
Restaurant Marketing Workshop