PracticeGraph
Appendix A · real specimens, measured locally
“You spent 57 hours last month watching an agent work.”
Agent runs are cheap to start and quiet to wait on. The minutes between your prompt and its answer feel free — they aren't. PracticeGraph adds them up, so the second currency of AI-assisted work shows up next to the first.
“The fix was ‘quick’ at 11pm. It's 1:47 now and you're re-prompting.”
Late-night sessions drift. Failures at 1am get answered with faster re-prompts and less reading than failures at 1pm. The spiral isn't a character flaw — it's a pattern with a timestamp.
“Nobody plans a six-hour session. It compacts three times and becomes one.”
Long sessions rarely feel long from inside — the context window compacts, the thread continues, the afternoon disappears. PracticeGraph counts the compactions, because that's where marathons hide.
“The answer arrives. Your follow-up leaves in under thirty seconds. Did you read it?”
Fast follow-ups after untooled answers usually mean the answer wasn't read — it was skimmed for the shape of done. The instrument can't see your eyes, but it can see the clock, and the clock is honest.
“A long agent run ends. ‘Looks good’ — in four seconds.”
The longer an agent works alone, the stronger the pull to wave the result through. Approval moments — where a long tooled stretch ends and you reply fast and short — are countable. Counting them is how you notice the habit before it costs you.
Named, not judged.
Every pattern above is an observation computed on your machine and seen by no one else. PracticeGraph never diagnoses, never gamifies, and sends nothing anywhere. The conclusions are yours.