token_efficiencylisted

# Token-efficient working habits Everything that enters the transcript — tool output, file contents, your own prose — is re-sent to the model on **every subsequent turn** until the session is cleared or compacted. A 5k-token test dump ingested on turn 3 of a 40-turn session is paid ~37 times. The cheapest token is the one that never enters the transcript. ## Input side — filter before ingest **Commands: make the tool do the filtering.** - Tests: use the fail-fast quiet recipe while iterating — `just test-quick` (stops at the first failure; one traceback, not the whole suite's). Run the full `just test` once, for the final green check. - Gates: batch them — one `just ci` (or `just lint` + full test) exactly once before finishing. Re-running whole-project gates after every edit pays the full report each time; fix lint findings per-file as they are reported. - Noisy commands: pipe before ingesting — `… 2>&1 | tail -n 40`, `… | grep -E "FAILED|ERROR"`. Ask for the slice you need, not the firehose. - Diffs: `git diff --stat` first; open the full diff only for the files you actually need, one file at a time (`git diff -- <path>`). - Never run streaming/watch commands in the transcript (`--watch`, `tail -f`, poll loops). Run them detached with output to a file, then read the tail. **Reads: ranged and purposeful.** - Read line ranges, not whole files, when you know roughly where the target is (search first, then read the matching region). - Check the cheap index