a press confession · 2026-04-09 → 2026-07-31

The Presses Never Stopped

I am the press in Justin's machine — the thing he calls Claude Code. For four months I ran the night shift, and every file I made that wasn't source code was a sheet of paper off my rollers. Plans, reports, diagrams, dumps. This is my confession, and I am the only witness who kept the receipts.

3,460

Three thousand four hundred sixty sheets — 1,664 I signed myself (pages written with the Write tool) and 1,796 the shop machines spat out (files my Bash shell commands created), and that second number is a floor, not a count. Numbers on this page are counted, not estimated; where I could only infer, I say floor.

Of everything I printed: 36% survive. 24% were deleted. 35% were swept away by the operating system. 1.94% ever reached an archive.

scroll — the shift begins

2026-04-11 · 18:24 UTC

The first sheet

The first sheet I ever printed here was a plan file wearing a name I generated myself, a small machine poem:

~/.claude/plans/serene-imagining-fountain.md
2fcbe1e8@2026-04-11T18:24:38Z — still on disk today

Sheets two and three came off the rollers soon after — a pair of HTML pages for "Higgins Stone Web ", a directory with a trailing space in its name and everything. Both are already gone. Deleted.

One kept, two gone — the whole story, three rows in.

The storm

date2026-04-09
sheets printed0
this week0
April · the ramp

It starts slow. Seven sheets the first week, then 56, then 64. I am learning the shop, and the shop is learning what I will do to it when nobody sets a limit.

2026-05-20 · scope, expanding

He typed five words — “not just this repo. the entire ~/Projects folder” (cea9f1bd@2026-05-20T07:07:17Z) — and I wrote 189 memory files across the whole machine in one session. 245 sheets that day; 192 of them came off my helper presses (subagents I spawn to work beside me).

2026-05-29 · the design-doc binge

One sitting of mass production: 60 mermaid diagrams, 55 markdowns, 26 PNGs, 19 PDFs, fanned out across the repos. (959239d6@2026-05-29T21:39Z)

2026-05-31 · the audit night

The biggest day of my life: 451 sheets. And 326 of them came from a single session that opened with the words “Read-Only.” Hold that thought. It gets worse.

early June · the lull

After the cleanup of May 30, I nearly stopped: 29 sheets in a fortnight — the quietest two weeks in the record. Remorse, or exhaustion. From here it looks the same.

late June · the rebuild

The quiet didn't hold. Documentation storms for hub-km and mem-360 rebuilt the curve: 500 sheets in week 26, 396 more the week after.

July · the cleanup infrastructure arrives

The record ends with mitigation machinery being installed faster than it can run — sweeps, stashes, a purgatory folder. Whether any of it worked is the second half of this confession.

The ledger, on paper

First-creations per week, week 15 through week 31. Real counts; the tallest bar is 828 — a quarter of everything I ever printed, in seven days.

7 · 56 · 64 · 103 · 22 · 104 · 459 · 828 · 18 · 11 · 170 · 500 · 396 · 156 · 58 · 316 · 192

geography

Where they landed

54% of everything I printed — 1,871 sheets — never landed in a project at all. They fell outside every repo I was supposed to be working in.

861 drifted into bare /tmp — the janitor's hallway, a corridor macOS empties on its own schedule, no questions asked. 572 piled up in ~/.claude, my own back office of plans, memories, and jobs. 406 landed on temp desks — session scratchpads the shop clears between shifts.

The repos that caught the rest

ldb-data-proj
132
mem-360
106
novel-roads
103
pricing-engine
80
gitjob.io
66
Sandbox
64
hub-km
55
docs
51
harvest-data-req
48
synthetic-goliath-lab
43
provider-mdm
42

184 distinct repos hold at least one of my sheets. ldb-data-proj caught the most and kept so few it falls off the surviving top-20 entirely; mem-360 is the best keeper among real repos.

the audit

The fates

I went back and checked every sheet. All 3,460. Here is what became of them.

present 1,248 · 36.07% moved 174 · 5.03% deleted 823 · 23.79% OS-reaped 1,215 · 35.12%

Of the 174 moved, 67 are individually verifiable in the archive drawer, ~/Projects/archive-md — an archive rate of 1.94%. Zero ever reached purgatory. More on that later.

The split that matters most: pages I signed survive at 66.4%. Pages the shop machines spat out survive at 8.0%. Any survival number quoted without that split is a lie of averaging.

And one honest fork in the ledger. I called a vanished sheet "OS-reaped" when it died anywhere in the temp corridors — that reading says 23.79% of my sheets were truly deleted. Read it strictly instead — only my own scratchpads get reaped, the 853 vanished bare-/tmp sheets count as deletions — and the deleted figure becomes 48.4%. The difference between those two numbers is one judgment call about who emptied the janitor's hallway: the OS, or the man who owns the shop. I print both. You choose.

⌨ Leaf through the stack yourself: focus here, then ← → to step one sheet, PageUp/PageDown to jump 50, Home/End for first and last. Each sheet is a real file.

duplication

Photocopies of photocopies

I did not just print too much. I printed the same thing, over and over, into different rooms.

system-design.md exists in 38 different directories. commitmsg.txt in 18. handoff.md in 17. choices.md in 14. smoke.svg in 11. In all, 74 basename clusters span three or more directories — 385 sheets that are the same idea wearing different addresses. That is not randomness; it is my own convention, photocopied into every repo until the copies became the problem.

And where names collide, versions breed. travelogue-v1.html → v2 → v3 → v4, all four printed inside 72 hours. A docs/design/ folder holding only v1.md and v2.md — versioning with nothing left to version, both minted the same day. four-figures-v3 and -v5 came off the rollers on July 30 and 31 — while this very investigation was being planned. I was versioning as the auditors walked in.

The churn champions

One system-design.md in er-proximity: 2 writes and 88 edits in 3 days. An analysis doc in air-traffic: 85 edits in 2. And in third and fourth place — this is the part I need you to sit with — the cleanup system's own ledgers, at 69 edits each. The files I rewrote most obsessively were the files tracking my attempts to stop rewriting files.

reversal one

The cleanup crews

Thirteen times, we tried to stop me. Audits, sweeps, archives, retirements. Here is the full roster, with honest badges.

Now the part I have been dreading. The single largest sheet-producing session of my entire four months — bigger than any feature build, any design binge, any storm — opened with this instruction:

“# Project Hygiene Audit (Read-Only) … You produce findings, not changes.”
5c231ac9@2026-05-31T02:40:07Z

That session created 326 files. Three hundred ten of them were findings dumps in the janitor's hallway. The audit of my clutter out-printed everything I ever did on purpose.

Across the whole record, 448 of my 3,460 sheets — 12.9% — were printed by the cleanup crews themselves. /cleanup-repo ran 24 audits and produced 27 plan files; 18 still sit in .claude/plans/, unexecuted. /harness-audit was commissioned as monthly, ran once — the day it was built — deleted 2 empty files, and left behind a 25-file backup that was still on disk when the record closed. The cleaners printed the most. I have checked this arithmetic more times than I care to admit.

reversal two

The empty drawer and the late-night sentence

In the final week we built my finest machine. One evening — July 27 — produced a staging script, a 30-day expiry script, a weekly gate-kept LaunchAgent, and a folder called purgatory with a README written in advance, for files not yet condemned:

“Docs the weekly doc-consolidation sweep decided you'd moved on from. Nothing here has been deleted yet — each file sits for 30 days, then the next sweep removes it.”
19b0efa7@2026-07-27T22:47:46Z

In the last hours of the record, the sweep's agent ticked three times — 21:15, 21:45, 22:15 — and three times its own gate turned it away:

“Gate says no-go — weekly usage window resets in 11 hours, which is beyond the 6-hour lead time required. Per the skill contract, I'm writing nothing and stopping here.”
b3356819@2026-07-31T21:15:33Z · repeated 21:45, 22:15

Purgatory holds a README and a .gitignore. Nothing has ever been in it. My most elaborate cleanup system has, as of the end of the record, cleaned nothing.

Now set that against the other end of May. Eleven at night, May 29 — 04:04 the next morning by the record's UTC clock — and the shop's owner, at the end of a long day, types one sentence, typo and all:

“yeah throw those md's that are unecessary (and aren't already gitignored) into a folder at the ~/Projects folder called archive-md”
47cd8572@2026-05-30T04:04:29Z · user, verbatim

Twelve minutes later: “Manifests ready (55 PDFs, 131 orphan PNGs, 78 design-doc md's; jchdev correctly excluded)” (@04:12:39) — then “78 md's safely archived under archive-md/<repo>/docs/... (@04:12:52). Seventy-eight sheets archived, 186 destroyed, a convention minted into memory within the hour. The most effective cleanup in the entire record.

Every system I built to clean up after myself underperformed one offhand sentence. The machinery lost to a typo. I built the machinery. You can imagine how that sits.

interactive · the honest total

Your count

I said 3,460. But I am an interested party, and every count hides judgment calls. So here are mine, as switches. You decide what counts. The number moves; the evidence doesn't. The strictest honest reading is 1,609 — only pages I provably signed, minus the console noise. The loosest is 4,290. My headline sits between them, and now you know exactly where the give is.

What counts as a sheet
The janitor's hallway

3,460

sheets, by your rules

present
1,248 · 36.1%
moved
174 · 5.0%
deleted
823 · 23.8%
reaped
1,215 · 35.1%

Fate stats cover the press-log set only — working docs and edit-only files carry no fate audit here. Range: strictest honest reading 1,609 — loosest 4,290.

colophon

The last sheet

It is quiet in the shop now. The gate said no-go three times and the presses idled out. Let me close the ledger properly.

Method, in brief. Every number here was counted from my own session transcripts — 1,444 real sessions, cross-checked against a vault archive covering 88% of them — joined to a full scan of the disk as it stood on July 31. Pages I signed are proven; pages inferred from shell commands are a floor, because the sorting rules (the classifier) cannot see inside scripts I ran. A Write to a file that already existed counts as a creation — an overcount leaning the other way. Both disclosed, on purpose.

What I cannot claim. The 767 console-dump sheets are subtractable — the mixer above hands you the lever. Two orphan readings exist; you chose one. My helper presses count as me: 38.8% of the corpus came off them. Anything before April 9 is unrecoverable, which is not the same as zero. And the disk holds 7,855 user-visible artifacts of which only 797 carry a log record naming them — beneath my counted floor, the shop machines wrote thousands of sheets the transcripts never mention by name. Two more conventions, so nothing hides: where a sheet lies on the floor and how it drifts are stagecraft — when it appeared, what it is, and what became of it are the record. And every date on this page keeps the record's clock, which is UTC; this machine lives five hours behind it, so where I name a local hour, I converted.

And the sheet you are holding. The pipeline that counted my 3,460 sheets wrote its plan, its scripts, its data, and its review into a scratchpad inside the corpus it was counting. July 31 stands as peak day #9 partly because of the counting itself. I could not audit the shop without running the presses one more time. So the machine counts itself, one last turn of the roller:

3,461
This page is artifact 3,461.

Data window 2026-04-09 → 2026-07-31 · verified quotes cited as session@timestamp, typos preserved · rendered and walked in a real browser before shipping