Airtable

Your assistant can list every base you have access to, then read the full schema of one: the tables, the fields, the field types, and how they are configured. You get that as plain description instead of clicking through twenty views. Nothing is changed by asking.

What you end up with. A written map of the base: which tables exist, what each field is for, and which ones are formulas or links to other tables rather than typed data.

Start with it bringing you the information and you deciding.

Never without. Never treat the schema description as a description of the data quality. It tells you the shape of the table, not whether anyone has been filling it in.

Built on: List bases / Get Base Schema

Your assistant can list records with a filter and a sort applied at the query, so you can ask for the open items owned by one person, sorted by date, and read the answer. It can then pull one specific record in full when you want the detail. Building a filtered view for a question you will ask once is the work this removes.

What you end up with. An answer to a one-off question about your data, without a new view cluttering the base afterward.

Start with it bringing you the information and you deciding.

Never without. Never assume it read the whole table. Record listing is paginated, so ask how many records it looked at before you quote a count or a total to anyone.

Built on: List records / Get Record

There is an action built exactly for this. It reads the table schema first, works out which of your fields the description maps to, and creates the record with type conversion switched on. You can also hand it structured values and have it create up to ten records at once. The type conversion is the part to watch, because it will make a reasonable guess rather than refuse an ambiguous value.

What you end up with. New rows in the right table with the fields populated, created from how you would have said it out loud.

Start with it doing the work and checking with you.

Never without. Never let it create records in a table other people report from without reading back what it wrote. Type conversion turns an ambiguous date or a near-miss select value into a confident wrong answer.

Built on: Create Record From Natural Language / Create records / Get Base Schema

It can update up to ten records in a call, changing only the fields you name and leaving the rest alone. Airtable also exposes a replace-style update that clears every field you did not specify, which is a different and much sharper tool. Ask for the selective update by name, and be aware the batch is not atomic, so a partial failure leaves some rows changed and some not.

What you end up with. A backfilled set of rows, plus a list of the ones it could not resolve and left for you.

Start with it doing the work and checking with you.

Never without. Never run a backfill without first asking it to show you the proposed change for the first few rows. And say explicitly that you want the selective update, not the PUT variant, because the PUT variant wipes every field you did not mention.

Built on: List records / Update record / Update multiple records

Airtable's own wording is worth reading before you ask: record deletion is permanent. It can remove one record or up to ten in a batch. The safe shape is two steps. First have it list the records matching your rule so you can look at them. Then, separately, tell it to delete that specific set.

What you end up with. The agreed rows gone, and a written list of exactly what was removed so it can be re-entered if the rule turned out to be wrong.

Start with it doing the work and checking with you.

Never without. Never delete from a filter description. Have it list the matching records first, read that list yourself, then approve deletion of that exact set, because Airtable removes records permanently and there is no undo through the connector.

Built on: List records / Delete Record / Delete multiple records

Comments on a record are readable and writable. Your assistant can read the existing thread on a record, then post a status note or a reply into it, including mentioning a collaborator. That puts the update where the work is rather than in a channel nobody scrolls back through.

What you end up with. A record whose comment thread carries the current status, with the right person mentioned on it.

Start with it doing the work and checking with you.

Never without. Never let it mention a collaborator without you seeing the text first. A comment mention sends a notification to a real person and you own how that reads.

Built on: List Comments / Create Comment / Get Record

It can read the attachment field on a record and return the metadata: file names, types, sizes, and the URLs. Combined with listing records, that means you can ask which rows have an attachment matching what you remember, rather than opening rows one at a time. It reads only, so it cannot lose the file while looking for it.

What you end up with. The specific record and the file you were trying to find, with its name and type confirmed.

Start with it bringing you the information and you deciding.

Never without. Never paste an attachment URL into a shared document without checking who that link works for. It points at a file governed by the base's own sharing, not by the document you paste it into.

Built on: List records / Get Attachment Info

It can create a table in an existing base with the fields you describe, add fields to a table later, rename a field, and rename or redescribe the table itself. It can also create a whole new base in a workspace with its tables and fields defined up front. Structure work is worth doing conversationally, because saying what you want to track is faster than choosing field types one at a time.

What you end up with. A table with named, typed fields ready to take data, built from a description of what you need to track.

Start with it doing the work and checking with you.

Never without. Never let it add or rename fields in a table that automations, formulas, or a synced view already depend on, without checking the schema first for what points at that field.

Built on: Create table / Create Field / Update Field / Update Table / Create base / Get Base Schema