ipeterpetrus Edit discipline
Measured on 1,316 Claude Code transcripts, 237,541 assistant turns (Aug 2026). Cost is carry = size x turns remaining: the transcript is replayed as cache-read input on every turn, so session input grows O(N^2). Producing a thing is cheap. Leaving it resident is not.
Overwriting a file that already exists
| condition | do |
|---|---|
| byte-identical to what is on disk | do not write — say it is already correct |
| under ~25% of the file changes | anchored Edit |
| over ~40% changes | full rewrite, and say why |
| 25-40% | either — take the one less likely to misapply |
Changed fraction, not block count. "<=3 change blocks -> Edit" loses ~79% of the available saving on the full corpus, because one merged block can span most of a file; the fraction rule keeps 84% of it held out over 20 session-level splits and was never negative. 20.8% of overwrites were byte-identical to disk — free to skip. Edit calls failed 6.2% of the time against Write's 12.2%, so anchoring is not the riskier path.
Where the tokens actually are
421-session cohort; a later 1,080-file run put Bash at 45.1% and Read at 19.8%, so quote a
share with its corpus. tools/carry.py prints attachments per type; this table sums them.
| source | share of carry |
|---|---|
| Bash call + result | 42.5% |
| Read results | 21.3% |
| injected banners (skill list, hooks, reminders) | 15.3% |
| Write + Edit calls | 9.5% |
| assistant prose | 5.6% |
| human prompts | 4.2% |
| other tools | 1.5% |
Everything in the section above is worth ~0.1% of a session. This half is worth more: read a range, not a whole file (mean Read result 22.8 kB, 25x a Bash result); ask Bash for the answer, not the log; and end the session — halving N halves carry, which no rule on this page can do.
The instruction itself
Prefix the request with one sentence, in the language you actually prompt in. Measured
against no instruction, on billed output tokens, 16 pairs each, facts retained 100%:
Jawab sesingkat mungkin, tanpa mengurangi isi teknisnya. −22.7% (14/16, p = 0.0001);
Answer as briefly as possible, without reducing the technical content. −19.6% (12/16,
p = 0.0070, pre-registered). A 4,664-character always-on block saying the same thing added
−0.8% over the sentence (p = 0.86). Do not ship the block.
The minimal-change sentence ("make the change as small as possible; add no unrequested
abstraction") failed two pre-registered tests; no number is quoted for it. Sentences were
measured as a prompt prefix, blocks as a SessionStart banner — never as a skill body like
this one. Protocols, intervals and every retraction: docs/FINDINGS.md section 9.