The model is not the system

The failure that looks like a prompt problem

You type a clear request. The model answers something adjacent. You clarify. It forgets the original goal. You rephrase. Forty minutes later you have a fluent document you do not trust.

That is not a prompt-craft failure. It is a missing operating system.

A conversationalist starts every turn from whatever is in the box. A system starts every turn from defaults you already installed: how it must behave, which workspace it is in, which files it may see, and how you will check the work.

The previous workaround was: paste Custom Instructions, start a new chat, and attach a charter. That still beats chatting with no rules. It is no longer the product.

How every answer should behave

Global Custom Instructions, ~/.codex/AGENTS.md, or user CLAUDE.md.

How this body of work should behave

Project instructions plus CHARTER.md, or repo AGENTS.md or CLAUDE.md.

A procedure you reuse

A Custom GPT or a SKILL.md. Not another paragraph in the global file.

What is true only today

The turn prompt. Not Memory. Not the operating system.

If you put this week’s OKRs in Custom Instructions, next month the model will argue with a ghost.

OpenAI Projects are workspaces. They hold chats, files, and instructions. Inside a Project, project instructions override Custom Instructions. You can choose default memory or project-only memory. As of 2026-08-14, you can switch that on an unshared Project. Shared Projects stay sealed.

If your current habit is to open a new chat and paste a charter at the top, you are running last year’s workaround on this year’s product. Create the Project. Upload the charter as a file. Put the job description in Project instructions.

1. 01

   Brief

   One-sentence objective, definition of done, constraints, sources in play, and what to avoid.

2. 02

   Plan

   The model restates the job, names the artifact, and names risks. You approve or correct. Do not skip this on anything that carries your name.

3. 03

   Build

   Use two passes when the artifact is prose. Use plan, implementation, and tests when the artifact is code.

4. 04

   Verify

   Separate facts, assumptions, and time-sensitive numbers. Or use tests, a diff, and a human gate.

That loop remains valid. Its previous install surface does not.

1. 1.Open Custom Instructions. Keep behavior, structure, and constraints. Delete anything that will be false next month.
2. 2.Create one ChatGPT Project named after an outcome, not a department. Pick a memory mode on purpose.
3. 3.Upload a one-page charter: objective, definition of done, constraints, and what to avoid.
4. 4.Run one real task through Brief → Plan. Stop there if that is all you have time for. Approving a plan is already a power-user move.

If the artifact is a repository, stop. Do not ask ChatGPT to invent a patch from a pasted file. That work belongs in a harness with AGENTS.md or CLAUDE.md, a verification skill, and a human who owns merge.