Custom instructions are standing directions to the model — how to answer you, what to assume, what to avoid — applied automatically so you do not retype them in every conversation. Secret Chat AI has two levels of them: global, and per project. This is one of the features of Secret Chat AI, the AI aggregator.
Global instructions
Set once in Settings, applied to all your conversations, on every model. This is the place for things that are true of you rather than of a particular task:
- How you want answers formatted — short and direct, or thorough with the reasoning shown.
- What you do, when it changes what a good answer looks like: "I'm a paediatric nurse", "I write Go, not Python", "assume US federal law unless I say otherwise".
- The tone you want, and the habits you do not: no preamble, no apologising, no restating the question back at you.
- The language you would rather be answered in.
Project instructions
Every project can carry its own instructions, applied to every chat inside it. This is the place for the context of a specific body of work — the case you are arguing, the codebase you are in, the voice a particular client's copy has to be written in, the facts the model should treat as given.
The practical effect is that a new chat in that project starts already briefed. You open it and ask your question, rather than spending the first two messages re-establishing where you left off.
Writing instructions that work
- Be concrete. "Be helpful" changes nothing. "Answer in under 200 words unless I ask for detail, and lead with the conclusion" changes every answer you get.
- Say what to avoid, not only what to do. Negative instructions are often the ones that make the difference — no bullet-point lists, no hedging, no suggesting I consult a professional when I have told you I am one.
- Keep global instructions genuinely global. Anything that only applies to one piece of work belongs in that project, otherwise it quietly distorts unrelated conversations.
- Revisit them. Instructions written six months ago are still shaping today's answers, which is easy to forget when a model starts behaving oddly.
One thing to keep in mind about privacy
Instructions are sent to the model along with your message — that is how they take effect. So they are subject to exactly the same rule as everything else you type: Secret Chat AI removes you from your queries, but not the data from them. An instruction that names your employer, your client, or your medical condition puts that text in front of the model on every message, not just once. Write instructions in terms of roles and requirements rather than identities: "I advise clients on employment disputes" does the work that a client's name would, without the exposure.
Both kinds of instructions are stored in your own browser alongside your chats, and project instructions travel with your exported data.