Canvas Tool Breaks Prompt Contract – Risk for GPT Adoption in Production Apps

ChatGPT
1

Hi OpenAI Team,

I’ve been building real SaaS apps with GPT-4, including a Telegram bot with Google Calendar integration.
However, I’ve repeatedly encountered severe regressions when using GPT with Canvas or memory tools:

:red_circle: It overwrites entire files, removes essential code, ignores uploaded context, and breaks strict prompt contracts.

This is not a single bug. It’s a systematic issue that makes GPT unsuitable for production-level work unless addressed.

:white_check_mark: I’ve documented the full issue + impact + suggestions in the PDF below:
[Upload the PDF “GPT_StrictMode_Feedback_Evgenii.pdf” here]

Please escalate this to the product team — I’m not looking for a support reply. I care about the future of GPT as a serious tool for real-world devs.

Thanks in advance.
:receipt: Feedback Report: GPT Prompt Contract Violation and Product Risk

Project: Telegram bot with Google Calendar integration
Primary file: telegram_bot_calendar_integrated.py
Original size: ~12 KB
Model: GPT-4 with canvas / ChatGPT Plus
Dates: May 18–19, 2025
Mode of use: strict prompt-defined behavior (clearly documented)

:red_circle: Issue Summary

GPT repeatedly violated the core instructions of a professional-grade prompt contract:

  • :cross_mark: Overwrote full source files with simplified versions, losing functions and handlers
  • :cross_mark: Removed essential imports (e.g. requests, StatesGroup), breaking runtime
  • :cross_mark: Operated on outdated canvas versions despite newly uploaded files
  • :cross_mark: Ignored instructions to apply non-destructive patches only (no replacements)
  • :cross_mark: Repeatedly returned untested or syntax-invalid .py files

:chart_decreasing: Impact

  • Severe loss of trust in GPT as an assistant
  • Time wasted rolling back broken outputs — multiple days lost
  • Risk to production logic due to corrupted file versions
  • Emotional and operational exhaustion from unreliable iteration cycles

:white_check_mark: User Expectations

  1. GPT should behave as an engineer, not a suggestion generator:
    • No destructive edits
    • No loss of logic or structure
    • Syntax and dependency checks before delivery
  2. Prompt contract is mandatory, not optional
  3. Ability to disable canvas or memory tools that cause regression

:light_bulb: Strategic Risk

From a consultant with 20+ years in brand and product strategy:

“Companies don’t fall because of market share or speed — they fall because they stop listening.”

Think:

  • Enron: collapsed by ignoring internal reality
  • Pan Am: vanished by failing to adapt to feedback
  • BlackBerry, Nokia: destroyed by arrogance, not competitors

OpenAI is becoming critical infrastructure. This feedback is a signal, not a complaint.

:rocket: Suggestions

  • Add a strict-mode toggle at GPT level (no canvas, no deletions, only diff/patch)
  • Validate file length + structure before applying edits
  • Show file diff before confirming a replacement
  • Improve session persistence and prompt-following over iterations
1 Like