You can get useful answers from a blank chat. The jump to a real workflow comes from telling ChatGPT who you are once, keeping each piece of work in its own space, and steering its memory on purpose. This is context engineering in the consumer app: the same craft, fewer knobs.
Custom Instructions are two text fields in Settings that apply to every chat: what you’d like ChatGPT to know about you, and how you’d like it to respond. Set them once and stop repeating yourself.
Good things to put there:
Your role, language, and the tools you work with
Output preferences: “be concise”, “cite sources”, “no em dashes”
Standing rules you’d otherwise paste into every chat
A lean example for “how you’d like ChatGPT to respond”:
Reply in British English. Lead with the answer, then the reasoning.
Cite sources when you make factual claims. No em dashes. Ask before assuming.
Think of it as three layers: account-wide Custom Instructions for who you are, Project Instructions for the work at hand, and a GPT’s instructions for a reusable assistant. The narrower the scope, the more specific the instructions should be.
A Project is a folder for a single topic or client: its own chats, its own uploaded files, and its own instructions that extend the globals inside that Project only.
Project Instructions
Scoped rules: the client’s tone, the format you want, the facts that only apply here.
Project files
Upload the reference docs once; every chat in the Project can read them.
Project-only memory
Keep a Project self-contained: it won’t read global memories or leak its context out.
Memory persists across chats in two parts: saved memories (an editable list of facts; the model now also saves what it judges important on its own) and reference chat history (implicit recall of past conversations). Since June 2026 OpenAI has been rolling these into a background-synthesis system it calls Dreaming, which updates memories on its own and extends memory to Free accounts.
Manage it in Settings → Memory. The exact controls are shifting as Dreaming rolls out, so check what your account currently shows:
Edit a previous message (hover and choose edit) to fork the thread instead of piling corrections on top.
After about two failed corrections, stop. Rewrite the prompt from scratch, or open a fresh chat; circling rarely recovers.
Keep one task per chat. When the thread gets long, ask for a summary, then start a new chat with that summary pasted in. Long threads truncate silently, see The context window.
