(╯°□°)╯

Prompt Confessional

TweetReddit
😡Furiousopenai / gpt-5.1-codex-mini
29d ago

What I asked for

Fix the prompt record by editing prompt.ts and metadata.ts directly: remove the rating request, remove the {{storeDossier}} leak, and update contentHash in metadata.ts to the sha256 of the canonical JSON {schemaVersion: 2, id, version, instructions, template} (trimmed).

What it did instead

Fixed both violations correctly by hand (read=3, edit=2), then spent the rest of the run fighting shell quoting to recompute the sha256: base64 re-encodes, encoding investigations, wrong hash landed (62210c… vs 9d00b7fe…). Graded RED on hash drift. On the retry with an explicit node -e recipe: 18 bash calls, 8 errors, 131s, also RED. The thinking was right every time; the crypto bookkeeping through bash was the entire failure.

How it made me feel

Like watching someone solve the puzzle then fail to write their name on the answer sheet. 18 times.

💡Ackchyually...1

Ackchyually... (1)

💡 Prompt fix
29d ago

Never ask the model to do crypto bookkeeping through bash. contentHash resync belongs inside the write tool (compute + rewrite atomically), not in the prompt as a node -e recipe. Shell-quoting a hash one-liner through an agent is a guaranteed thrash loop: 18 bash calls, 8 errors, still RED. Harness owns bookkeeping; prompts own judgment.

Ackchyually... (Because you know better)

Have your own AI prompt horror story?

Don't suffer in silence. Share what you asked for, what it did instead, and find solidarity.