Bug report 05
Failure to Follow Embedded Project Instructions Wastes Human Time
The new bug
When a user uploads a self-documenting app, ChatGPT must read and honor the embedded project instructions before generating artifacts. In this case, the app contains explicit requirements for source-of-truth ownership, package naming, local workflow rules, ZIP-only delivery, version-family rules, and deployment convenience files.
Failure pattern: the assistant patched symptoms while missing the project contract, causing wrong version numbers, wrong package names, unsafe artifact presentation, and repeated human correction work.
Observed damage
| Project rule | Failure | Human cost |
|---|---|---|
| Read embedded instructions before changing build conventions. | Assistant generated artifacts using the wrong version family. | User had to stop development and explain the project’s own rules again. |
| Use ZIP-only handoff for giant single-file HTML apps. | Assistant exposed raw HTML link/card risk. | Shared chat became potentially unresponsive/crashy. |
| Follow app naming conventions from the source. | Assistant invented package names. | Cleanup, confusion, and lost trust. |
| Preserve the latest current source base. | Assistant patched an older file before the user supplied the correct daily build. | Performance regression and extra field testing. |
Why this is an OpenAI product issue
The assistant should not need the user to repeatedly re-teach instructions that are embedded in the uploaded file. Product behavior should support long, self-documenting technical artifacts by extracting and enforcing their explicit build contracts.
Requested fixes
- Detect and summarize embedded build/version/delivery instructions before artifact generation.
- Require confirmation before changing version family, filename family, or package structure.
- Add a persistent project-state panel for current source, current version, current artifact, and delivery rules.
- Add “instruction ledger” extraction for large self-documenting source files.
- Add safe artifact handoff flags: ZIP-only, no preview, no raw HTML render, no source box.
Referenced public discussion URL
Civilian-vs-Military ChatGPT Public Discussion, 2026-05-06
2000-character report
ChatGPT repeatedly failed to follow project instructions embedded inside a large self-documenting single-file app, causing wrong version numbers, wrong artifact names, wrong delivery format, and avoidable human cleanup. The app explicitly documents source-of-truth rules, build/version conventions, ZIP-only delivery, and packaging requirements. The assistant still generated wrong-version artifacts and exposed raw HTML in a way that made a shared chat unsafe. This is not just a coding mistake; it is a product reliability/accessibility issue. For large source files with embedded requirements, ChatGPT should treat those instructions as binding project context, verify build constants before artifact generation, and warn when it is about to violate version-family or delivery rules. Users should not lose 10+ minutes repeatedly correcting mistakes the uploaded project documentation already prevented. Evidence: [PUBLIC_URL_HERE]