Reduce activity state fragmentation via prompt tuning
Analysis of Feb 9-10 journal data showed 43% false-new rate in
activity_state: the LLM was marking coding as "new" on consecutive
segments because the description changed (different files, different
subtask), producing 33 separate coding records for a single day.
Key prompt changes:
- Define "continuing" as the default when same type is in previous state
- Narrow "new" to only when the type wasn't previously active or there
is a clear session boundary (different meeting, different participants)
- Add "prefer continuing" as the most prominent rule
- Scope same-type transitions to meetings/calls only, not coding/reading
- Clarify that description evolution is normal for continuing activities
- Rebalance examples from 3 new/1 continuing to 2 continuing/1 new
- Include "focused reading/review" as evidence of active engagement
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>