A scheduled prompt often starts without a person present to notice a missing attachment or a stale cache. Require each input to carry a source, collection timestamp, coverage interval, and retrieval status. An empty set from a successful query can mean no matching records. An empty set caused by a failed query means unknown. These states need different output. Set an age limit for data and a completeness threshold for expected partitions before generating a normal report. A model cannot repair a missing source by producing a plausible zero count.
Make this comfortable
Recurring prompts: validate fresh inputs and empty states
Operational case
At 08:15 the North Pier status endpoint returns zero held parcels, but its refresh timestamp is two days old. The previous verified snapshot contained 47 held records. Reporting 'all 47 released' would be false. The run marks the source stale and requests operator attention. On another day a fresh response returns an empty set with all expected partitions present; only then may the report state that no held parcels were found in that verified window. The evidence packet preserves which path produced each result.
Query status: success; snapshot age: 49 hours; limit: 3 hours.
Result count: 0; previous verified held count: 47.
Decision: stale source, state unknown; no release claim.
Fresh zero with complete partitions: verified empty set.
Failure to query: unavailable, never zero.Performance and operating cost
Checking P source partitions takes O(P) metadata reads before content analysis; a full reconciliation may add O(N) record work. The gate can delay a report, yet the alternative is a confidently wrong operational statement. Keep staleness limits tied to the business decision, not to a universal number. Log source IDs and timestamps without copying unnecessary customer fields into the prompt or telemetry.
Common Mistakes
- Do not translate a failed query into a zero result.
- Do not compare a stale snapshot with a fresh one as though both were current.
- Do not bury a missing partition in an average-quality score.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Retrieved evidence: reconcile versions and conflicting facts
- Selective answers: measure when to abstain
- Prompt telemetry: measure failures without copying private payloads
- Recurring prompts: define the trigger and stop condition
- Recurring prompts: bind local time and data windows
- Recurring prompts: survive retries without duplicate effects
- Recurring prompts: notify only on actionable changes
- Recurring prompts: renew authority for sensitive actions
- Recurring prompts: make every run replayable and auditable
- Project: build a North Pier recurring review
- Recurring prompt workflow decisions
prompt engineering
recurring workflows