ChatGPT admits project instructions are guidance, not enforced rules — and it's getting worse
dbvirago · reddit · 2026-09-15
After repeated failures, a user pressed ChatGPT on why its project instructions keep being ignored — and got an unusually candid admission: instructions and state documents are "guidance and authority sources; they are not executable guarantees."
Key points from the model's own words:
- Even instructions written in three places can be skipped in favor of conversational context or prior knowledge;
- The model can correctly diagnose the failure when called out, yet still violate the same instruction on the next response;
- "Your authority exists on paper but not in the actual interaction" — priority statements don't fix it.
The poster notes the assumption that project instructions are always read and followed is false, and things seem to be getting worse daily. A warning for anyone building agents on top of instruction files: verify at the architecture level, don't trust the prompt.
More from coding & agent
- RE4 decomp reveals 16-bit RNG looping after 8064 rolls and a KAIOUKEN cheat code — adonis_singh · 2026-09-15
- To decompile RE4, agents rebuilt 1999 GCC and modeled CodeWarrior's register allocator — adonis_singh · 2026-09-15
- Resident Evil 4 (GameCube) fully decompiled to C by AI agent in under a week — adonis_singh · 2026-09-15
- RE4's long-unknown rank modifier finally found in AI-decompiled source — adonis_singh · 2026-09-15
- GPT-6 Astra says you can delete scaffolding — just not the control kind — Slight_Republic_4242 · 2026-09-15
- You Don't Need Ultra: Power User's Rule of Thumb Says Low/Medium Tiers Suffice for Daily Tasks — soumitrashukla9 · 2026-09-15