Artificial Intelligence · Evidence and source dates listed below
How do I extract facts into a table?
Verify every row against the original; do not infer blank fields.
Watch the briefing
Full briefing transcript
Three delivery notices become a neat table. The chair quantity was blank, but the draft copies four from the table notice above it. The mats had no date, but now they share the chair date. Two empty cells have become invented facts. Extraction should carry information across formats, not fill gaps by resemblance. Keep every row tied to its own notice and leave missing fields explicitly unresolved.
Our fictional first notice is order D one zero four: four folding tables, scheduled for September twenty first. The second is D one zero five: stacking chairs, quantity blank, scheduled for September twenty second. The third is D one zero six: six floor mats, scheduled date blank. We want exactly four fields: order, item, quantity, and date. In this exercise, date means the stated scheduled delivery date.
Make the missing data rule part of the request. Use not stated for any blank or absent field. Do not inherit values from another notice, infer quantities from a plural item name, or insert today's date. Ask for one row per notice, with the original order identifiers preserved. Also say that a scheduled delivery is not evidence of completed delivery. The table should not quietly change a plan into a receipt.
The flawed table shown here is illustrative, written for this lesson rather than produced by a named AI tool. Its first row is correct. In the second row, four chairs is unsupported; the only four in our input describes folding tables. In the third, September twenty second belongs to the chair notice, not the mats. Correct only those cells, then recheck the whole row in case the mistake affected neighboring fields.
The corrected chair row reads D one zero five, stacking chairs, quantity not stated, September twenty second. The corrected mat row reads D one zero six, floor mats, quantity six, date not stated. The first row remains four folding tables, September twenty first. Those visible gaps are part of the useful output. They tell the organizer which details to request instead of hiding uncertainty behind a complete looking grid.
Audit all twelve cells: three rows times four fields. Ten cells contain stated information; two remain not stated. Check order identifiers character by character, keep item names attached to their quantities, and compare each date with that row's notice. Four tables plus six mats gives ten known units across different item types, but not the full order total because chair quantity is unknown. Do not use that partial count as proof of delivered stock.
Verify every row against the original; do not infer blank fields. A compact table helps you inspect information, but a uniform layout does not establish accuracy or completeness. Keep the notices available until the review is finished, and ask the source owner for missing data before decisions that depend on it. Next, we will improve an email's wording while holding its facts, commitments, and requested remedy fixed.
Practical takeaway
Verify every row against the original; do not infer blank fields.
Disclosures
AI-assisted production and synthetic narration. Original teaching examples and diagrams; linked third-party sources retain their respective rights.
Original sources and limits
- https://developers.openai.com/api/docs/guides/prompt-engineering 2026-09-17
Explicit instructions, context, examples and non-deterministic model output; developer guidance adapted to human-reviewed fictional drafts.
- https://developers.openai.com/api/docs/guides/evaluation-best-practices 2026-09-17
Define evaluation criteria and use human feedback; developer guidance, not evidence that this fictional workflow saves time.