Audit assistant habits from saved prompts and session transcripts
At a glance
A developer used Claude to analyze his own Claude Code history and turn evidenced patterns into reusable workflow instructions. This workflow is listed under Developer Workflows, Session Analysis, Personal Tools.
- Agents / tools
- Claude Code, Python, Text Search
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Give Claude access to retained ~/.claude/history.jsonl and session transcripts under ~/.claude/projects/. 2) Ask it to use text searches and Python to extract compaction events with four preceding and two following prompts, a time-stratified sample, model-switch history, and per-model message counts. Include prompts containing your delegation keyword where relevant. 3) Analyze compaction, delegation, prompting style, model choice, and planning separately. The author used five parallel analysts returning structured, evidence-backed findings. 4) Synthesize explicit triggers and recommended behaviors with supporting records. Ask for reusable skill instructions, then review which rules depend on your tools or habits before adopting them.
How it went
The author's history contained 27,939 prompts across 3,183 sessions. The synthesis produced 12 workflow behaviors.
Worth knowing
The author explicitly calls model-routing findings correlational and warns that his custom agent fleet and personal prompting style do not transfer wholesale.
