# A whole-file write dropped 57% of my CLAUDE.md. It looked healthy for three days. — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/fc11e7e4410e8008

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-26 17:44 ingested 2026-10-27 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xg3s17/a_wholefile_write_dropped_57_of_my_claudemd_it/)

## retained text



```
I run 28 projects out of one markdown vault. One file at the top holds every rule, every SOP and the project map. Every session starts there. Kind of like the Andrej Karpathy .MD that everyone is passing around.



One day, a Claude Code session did a whole-file write on it. The file went from 58,159 bytes to 24,982. It ends mid-word: "a missed spin-".



Eleven top-level sections were obliterated: The deployment SOP, versioning rules, folder structures, changelogs. The rule that says never rewrite a whole file was in that file.



I, nor Claude, noticed for three days. Three later edits added new content to what was left, so the file kept growing and looked healthy by size.  


A lucky audit of some of my system ended up raising the alarm and noticed a git backup had twice the info... The revision history had ended at the same mid-word; every copy in it was already short.



How do you catch errors like this day to day? Claude just isn't reliable when it comes to situations like this.
```



document fetched 2026-10-27 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xg3s17/a_wholefile_write_dropped_57_of_my_claudemd_it
