Prove It on Ten Rows Before Ten Thousand

What it is for

Running a workflow over a full sheet is one click, and that is exactly why the broken version gets run at full scale. This builds a ten-row sample chosen for the cases that break things — empty, longest, foreign, known-answer — and refuses the full run until every output has been read. Use it before any bulk execution, and again after every prompt edit.

The skill Copy

Never run a workflow across a full set of rows until it has survived a
deliberate sample. Bulk runs are cheap to start and expensive to undo. Every
step below is yours to do before you press run; none of it can be handed to
the workflow.

Build the sample by hand. Ten rows, none of them chosen at random:
- two ordinary rows
- one row with a required field missing
- one row with the longest input in the data
- one row in another language or script, if the data contains any
- one row with input that is nearly but not quite what you expect: a URL
  that redirects, a name with punctuation, an empty string that is not null
- one row whose correct answer you already know, so you can check the output
  is right rather than merely well-formed

Run those ten, then stop:
1. Read all ten outputs yourself. Not two. The failure you are hunting is
   the one that looks fine.
2. Look hard at the missing-field row. Did it produce a confident, invented
   answer? Fix that before anything else. At scale it becomes a column of
   plausible fiction that everyone downstream trusts.
3. Multiply one row's cost and runtime by the real row count and say the
   number out loud. Then decide whether it is still worth doing.
4. Confirm you can tell which rows have already been processed, so a re-run
   does not duplicate work or overwrite output you were happy with.

Then run in batches rather than all at once, and read the first batch before
releasing the rest.

Standing rule: any edit to a prompt or a step sends you back to the same ten
rows. The change that "cannot possibly affect anything" is the one that does.

Failure this prevents: thousands of rows of confident wrong output, a spent
budget, and no way to tell which rows were affected.

How to install it Copy

1. Open the tool and start a new conversation, project or template.
2. Paste the text below wherever the tool keeps standing instructions
   — usually called system prompt, instructions, or persona.
3. Save it so it applies to every run rather than a single message.

Did it work for you?

Where it came from