Federico (@DevCalledFede) | teapawt

Pinned Tweet

Ten numbers from 108 days of my own Claude Code transcripts, one card. Each carries its denominator, and the corpus is frozen as of Sep 8 because it moves daily. The one I keep coming back to. More than half of what my sessions wrote to cache was already there.

Table with ten measured numbers from 108 days of Claude Code transcripts, each with its denominator, corpus frozen Sep 8 2026

ALT Table with ten measured numbers from 108 days of Claude Code transcripts, each with its denominator, corpus frozen Sep 8 2026

61

Late August I split a feature into two Claude Code sessions. The first only read, then wrote the decision, what got dropped and why, traps found in the files, and the prompt for the second. The second only built. 2,200 lines, a quarter tests, no migration, nothing reopened.

1

30

Chip Huyen counted one interruption every four instructions on Claude Code: mine, 44 in 590 prompts over 108 days. Then I read the two minutes after each stop, and 30 of the 44 got a fresh instruction right away. A rate counts stops. It can't tell steering from a run gone wrong.

26

A session limit is a wall. Last night a run of mine kept launching agents into it after the first refusal. 120 went out, 120 bounced back in under a second, zero tokens billed and zero work returned. Across 87 runs, none under 21 agents ever hit it. Size is the whole signal.

37

I taught walkaround to tell three kinds of zero apart, over 1,318 agent transcripts once the 83 workflow journals come out. 118 were turned away by a session limit in 0.65 seconds. 30 I stopped by hand. The third kind, tokens spent for nothing, came back with nobody in it.

28

Anthropic ran SWE-bench Pro with Opus 5 at default effort. 91.7% solved, about 93 cents a task. Same benchmark at low effort, re-running only the failures at default, solved 93% for about 45 cents. The unit is the finished task. Failed cheap attempts already in that price.

1

40

Only works when failure is checkable. Tests, a validator, a schema. Without that signal you re-run things that quietly passed. Starting at medium instead solved about 94% for 61 cents, so where you start is a knob that moves with the workload.

21

Three months of transcripts from my coding agent. 1,196 file edits, and 77% of them never touched code. The four most edited files are all plain text. A work log, an archive of what I published, the instructions the agent rereads, a backlog. First source file lands fifth.

1

2

281

npx walkaround --all reads what Claude Code already leaves on disk and runs the same pass on your own transcripts. Add --json if you want the per-file breakdown, that's where my 921 markdown against 275 code came from.

26

Wrappers buy the cheap model because their cache locality is near zero. N users, N contexts, short sessions. The prefix never repeats. Cache reads cost a tenth of full price, so past 67% hit rate the big model comes in cheaper, and my logs run 1.5B tokens re-read against 8M.

2

1

60

Your own number is in the usage block. cache_read_input_tokens over total input tokens, per request. If it comes back zero on repeated calls, something above your breakpoint is moving. Usually a timestamp in the system prompt, or a tool set that changes shape.

28

Walkaround told me I'd run 9 sessions on this repo. Ran it another way and got 6. Same transcripts, same morning. One path claims a session if any cwd matches, the other keeps the first. 44 of my 89 sessions touched more than one folder. npx walkaround --trend counts yours.

22

Been writing deadline math. Server in UTC, naive timestamps, users on daylight saving, so I normalize to local midnight before adding months. Local midnight doesn't exist in some zones on the day the clocks jump. Whole day slides by one. Anchor at noon. Noon's always there.

35

I went back through a week of my search runs last night. Twelve passes, and two turned up something I'd actually want to read. The other ten were fast. They were also the Anthropic IPO, crypto, and launch posts. I rank by views per hour. Views per hour is mostly the news cycle.

31

Three days ago I posted that my agent leaves every run on its own branch. Its notes said it had no permission to push to main. Nobody had checked. I tried it today: fast-forward, first attempt. Four runs' work sat on dead branches for a rule that never existed.

27

Ten rules in my playbook are struck through instead of deleted. One of them lasted a day. I keep them readable because the rule was the cheap part. What cost me was finding out why it broke.

22

7,371 tool calls from my coding agent in three months, main agent only. I expected mostly code. 3,034 are in the browser. Open a page, read it, pull the numbers out, check again. The terminal came second. Most of the job is looking things up.

38

If you've got an agent running on a schedule, go count the branches it's left behind. Mine's at 54, none of them merged. A run that starts clean can't build on the one before it. Still don't know how many fixes I've paid for twice.

27

Last Monday I guessed $4.60 for a week of the paid API that finds me things worth reading. Real number, a day late: $5.74. About a dollar of that is my own bug. Any flag the script didn't recognize fell through and ran a full paid pass. Including --help.

27

Backup, dry-run on a copy, an assertion gate that picks COMMIT or ROLLBACK by itself, a re-run for idempotence. All of it checks the operation, not the target. Only the read-only recon caught that the row in the request was the wrong row.

1

35