# Check its work: the five-second card

Print it. Tape it next to the screen. Run it on every AI output before you send, ship, paste or upload. From Distru's No Bullshit AI Course, lesson 03-04. CC0 1.0.

Status 2026-09-17: the five checks are the manual version of the failure list in course lesson 01-02; the table at the bottom is the scripted version. No benchmark numbers. Keep a tally of what each check catches for a month and you will know which two to keep.

---

## The five checks

**1. Source.** Did I give it the thing it is talking about? If it summarized a bulletin, did I paste the bulletin? If not, the summary is from memory, and its memory has a cutoff date and no access to your state.

**2. Number.** Pick one number. Recount or recompute it by hand or in a spreadsheet. One line is enough to know whether to trust the rest. If the one you picked is wrong, check all of them or ask for the formula instead.

**3. Date.** Every date and every "this week / next Tuesday / Q3". AI drafts default to the wrong year more often than any other single mistake.

**4. Where did that come from?** Read once asking only that question. Any fact, name, price, rule or quantity you did not paste in is a guess. Find it in your material or cut it.

**5. Rerun once.** Send the same prompt again (or ask "what did you get wrong?"). If the two answers disagree on anything that matters, neither is trustworthy yet. Fix the prompt, not the answer.

## Then decide

- All five clean and it goes to a colleague: send.
- All five clean and it goes to a customer, a vendor, the state or the public: one human read, then send.
- Any check fails: fix the prompt (more material, clearer shape, the "say MISSING" line) and rerun. Do not hand-patch the output; the next run will make the same mistake.

## For the tech side (same five, made permanent)

| check | the manual version | the version a script runs every time |
|---|---|---|
| source | did I paste it? | log the prompt and its inputs next to every output (one JSON line per run) |
| number | recount one line | sum checks, row-count checks, schema validation on any structured output |
| date | read every date | regex for dates; reject any year outside the expected range |
| where from | read once for guesses | require quotes or row references in the output; fail if missing |
| rerun | send it twice | a small eval set: 10 to 20 fixed inputs with known right answers, rerun on every prompt or model change |

Spot-check rule of thumb for batches: check 10 rows of 200 at random. Zero wrong, ship. One wrong, check 30 more. Two wrong, stop and fix the prompt.

The agent version of this card, where a yes/no check runs before an AI is allowed to act, is course lesson 06-05.