How to Prompt an Airtable Schema from a Workflow
Airtable prompts fail by building a product. They add automations you did not ask for, rollups on fields you do not collect, and a sample row with a fake email. You paste it. Syncs fire. A real person gets a test message. The job is a schema from the workflow you run today, with only the tables and fields you named, and no automations beyond Inputs.
The matching generator is the Airtable Base Schema from a Workflow prompt. Browse related cards in the PromptDig library (Browse more prompts). When a filled run survives, share the version you actually use (Share a prompt).
Workflow first. Tables are nouns you already handle.
Describe the role and the path a piece of work takes: request in, triage, do, wait, done. Tools you already use (forms, Slack, email) stay named. Fields I actually have is the list of columns you already collect or will collect this week. Do not let the model add NPS, ARR, or a "sentiment" field because other bases have them.
Cadence is how often you touch the base. Pain today is the concrete mess (duplicate requests, no owner, lost attachments). Do not automate: the list of things that must stay human (client email, money movement, anything in DoNot).
Schema, then a configurable artifact without bonus robots
The schema map: tables, primary fields, links, field types that Airtable actually has, and views that match Cadence. If you named Requests, People, and Vendors, you do not get a fourth table called Analytics unless you asked. Linked records only between tables you named.
Configurable artifact: field list with types, a select's options only from Inputs, formulas only from fields you listed. If a formula needs a field you do not have, put it under Gaps, not as a magic rollup.
Sample row: fictional but not fake PII. Use clearly fake tokens like requester@example.invalid, or role names, not a scraped phone number. Do not invent a real company as a sample client.
Automations: only what DoNot allows and what Inputs described. If Inputs are silent on automations, write "none specified" instead of an "on form submit, email the CEO" flow.
Maintenance, version lock, and gaps instead of Interface theater
Maintenance cadence: who reviews selects, who archives done records, what not to sync. If you did not name Interfaces or a portal, do not design one.
Version-lock Airtable features only if you stated a plan or a feature. If unnamed, write unknown. Do not assume scripting, two-way sync, or a feature that may not be on your plan.
Gaps: missing unique ID, missing owner field, missing attachment policy, missing automation permission, missing PII rules. Five bullets. De-identify. This is ops design, not legal advice on retention.
The base should be smaller than the prompt dump you have seen on forums. That is the point.
A worked pass: three tables, no surprise email
Workflow: inbound request, triage, done. Fields: title, requester role, status, owner, due. Tools: Airtable, a form. Do not automate: anything that emails a client.
Tables: Requests, People if you named people, nothing called Analytics. Status select from the words you use (new, waiting, done), not a 12-stage enterprise funnel. Sample row: title "Replace hallway lamp," requester role "facilities," email example.invalid.
If the model adds an automation "when status is done, email requester," delete it. That was not in Inputs. Views: Mine, Waiting, Done this week, matching Cadence. Interfaces omitted. The base is a tray for the work you already do, not a product launch.
Fill the card, then run
Paste the real workflow, the fields you already collect, and what must never auto-send. Then run it and delete any automation the model smuggled in.
Workflow / role: [Workflow]
Tools I already use: [Tools]
Fields I actually have: [Fields]
Cadence: [Cadence]
Pain today: [Pain]
Do not automate: [DoNot]
A base is a filing system with types. It is not a second product. When a run refuses a fake email automation and keeps the sample row clean, share that filled card. That is the schema you can actually build this afternoon.