A report-writing prompt should receive verified metrics, a named audience, a decision question, and a required claim-to-evidence map. The assistant can turn numbers into readable prose, but it must not invent a causal explanation or omit a result merely because it complicates the headline. Distinguish measured change, plausible explanation, and proposed follow-up. Require the summary to mention excluded records and the period covered. If a headline depends on one region while another moves in the opposite direction, preserve that contrast. Review the final rendered summary against the aggregate table, not just against the prompt.
Make this comfortable
Report summaries: preserve contrary results and missing data
Operational case
A weekly parcel report shows a 6.00 percent late rate across 4,700 eligible shipments. North has 80 late out of 2,000, or 4.00 percent. South has 202 out of 2,700, about 7.48 percent. An assistant draft says service improved everywhere because the overall rate fell from a prior week. That claim is unsupported: the prior-week regional denominators have not been supplied, and 23 legacy records were excluded because their delivery status remains unresolved. A defensible summary states the current counts and exclusion, then asks for comparable prior-week cohorts before attributing an improvement.
Audience: operations lead; period: week W-39.
Checked: all=282/4,700=6.00%; north=80/2,000=4.00%.
Checked: south=202/2,700≈7.48%; unresolved statuses=23.
Allowed claim: south's observed rate exceeds north's in this cohort.
Blocked claim: a change caused improvement across every region.Performance and operating cost
Checking K narrative claims against M verified figures takes O(K+M) with an indexed metric map; manual adjudication of ambiguous claims is the expensive part. A claim ledger adds modest storage but makes later corrections traceable. Short summaries reduce reading time, yet compressing away the denominator or exception count can invert the decision. Evaluate summaries for numerical agreement, omitted contrary results, unsupported causal language, and missing caveats. A polished paragraph with one false claim is worse than a shorter statement of what remains unknown.
Common Mistakes
- Do not describe an observed difference as a proven cause.
- Do not omit a contrary region to make the headline cleaner.
- Do not compare weeks until the cohorts and exclusion rules match.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Evidence IDs: make generated claims auditable against supplied records
- Aggregation prompts: pin the denominator and recompute the rate
- Evaluation labels: adjudicate disagreement before scoring a release
- Dataset intake prompts: define a column before analyzing it
- Classification prompts: write the label boundary first
- Chart prompts: inspect axes, units, and missing series first
- Project: verify a shipment performance report
- Prompt engineering for data workflows
Continue with: Plain-language prompts: simplify without losing conditions.
Continue with: Conflicting studies: inspect design before averaging outcomes.
Continue with: Report prompts: reconcile every claim with the packet.
Continue with: Project: review a cold-chain temperature chart.
prompt engineering
data workflows