teardown · pycomet
103 capabilities installed, 90 never fired in 30 days
A month of session logs, mined for what actually ran. Almost nothing did.
01
The setup
pycomet built a read-only scanner for ~/.claude that inventories skills, subagents, MCP servers and hooks. It then mines local session transcripts for what fired in the last 30 days and uses a cheap Haiku call to sort capabilities into five areas.
02
What they found
103 installed capabilities, 90 of which never fired in 30 days. 25 pairs of literally identical skills, because one plugin was cached under two marketplace names. One forgotten skill loaded about 150 tokens into every session: about 11,000 tokens across 72 sessions, which the author labels a lower bound.
03
Our read
The useful idea is measuring use instead of guessing. The install-vs-use gap is invisible in config files and obvious once you count.
04
What to steal
Mine your own session logs for what actually ran, then delete the rest.
05
Caveats
These are one person's numbers for one month. Token figures are the author's own lower bounds.
Facts from https://dev.to/pycomet/i-audited-a-month-of-my-ai-coding-setup-87-of-it-was-dead-weight-pkp (2026-07-22). "Our read" is Agent Setups' analysis, not the author's words.
next teardown →
Broke up with Claude Code: open-weight models at about two cents a task