You've got a personalized system prompt — here's how to actually put it to work. Start with the shared idea, then open the tab for the assistant you use.
A system prompt is a standing instruction that shapes how an assistant answers — its tone, priorities, and blind spots — before you ask anything. Every type page gives you 3 ready-to-use prompts (coaching, careful execution, or critique); pick the one that matches what you want and copy it.
The flow is identical everywhere: copy your prompt, then hand it to the assistant either once for a single chat or as a standing instruction. Here is what that looks like in each of the three assistants.
For one conversation, open a new chat and paste the prompt before your first question. To make it permanent, open your account's personalization settings and paste it into the custom-instructions field — ChatGPT then applies it to future chats. For a middle ground, create a Project and paste the prompt into the Project's instructions so only related chats follow it.
For a single session, paste the prompt at the very top of a new conversation and ask your question below it. To reuse it, create a Gem: name it, paste the prompt into the Gem's instructions, and save. From then on, chatting with that Gem always applies your rules, and editing the Gem updates every later conversation.
Paste the prompt as the first message of a new conversation to shape just that chat. For a lasting setup, create a Project and add the prompt to the Project's instructions — every conversation started inside it follows the prompt automatically, while chats outside the Project stay unaffected. That makes it easy to switch your custom behavior on and off.
A system prompt is guidance, not a switch — most problems come down to small wording fixes you can make yourself.
In long conversations, early instructions gradually lose ground to recent messages. Restate the one rule that matters as a direct request — "answer in bullet points, five max" — or start a fresh chat so the prompt is the newest thing the assistant has read.
The prompt is plain text and you own it. Delete the line that pushes too hard, or soften absolutes like "always" and "never" into "prefer" — then paste the trimmed version. Small edits change the feel a lot.
Add one explicit line about length, such as "keep answers under 150 words unless I ask you to go deeper." Length rules work far better as numbers than as adjectives like "brief".
Two competing instruction sets produce muddled answers. Merge them instead: keep the preferences you already trust, then append only the lines from your type prompt that target your blind spots. One merged instruction beats two conflicting ones.
No prompt yet? Take the 3-minute test and get one built for your type.