Claude Code studies 27,939 personal prompt records

waizns asks Claude Code to analyse a personal prompt and session corpus rather than rely on remembered habits. The account identifies 27,939 prompts across 3,183 sessions and 210 days, with detailed tool and model transcripts available for inspection.

A Python extraction pass creates focused datasets: compaction events with surrounding prompts, orchestration requests, a time-stratified prompt sample, model-switch history and message counts. Five analyst agents separately examine compaction, delegation, prompting, model choice and planning. A sixth synthesises structured findings into a skill containing twelve behaviours, each connected to supporting telemetry and a trigger.

One result challenges the author’s assumption that compaction happens only at clean endpoints: the extracted records show it occurring inside long sessions, with substantial work continuing afterward. Another compares rarely invoked interactive planning with persistent written plan files. These observations describe this user’s corpus rather than a controlled test of which practice causes better output.

The completed artifact is an installable operating skill derived from an executed analysis. The article’s token and timing figures are self-reported; it does not independently demonstrate that applying the skill improves future behaviour. The useful research pattern is extracting inspectable event windows first, assigning a bounded question to each analyst and retaining evidence in the synthesis rather than turning personal patterns into universal productivity rules.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Blog · waizns