The basics
Every AI tool has a place where you can write down things like "I'm a lawyer in California, I work mostly on commercial leases, please always use plain English and skip the disclaimers." These are called custom instructions in ChatGPT, personalization or system prompt in Claude, and gems or personas in Gemini.
The AI reads your instructions before every single chat. So you don't have to say "I'm a lawyer" or "use plain English" again. It already knows.
What people put in their instructions
- Who they are. Job, industry, location, what they're working on right now.
- How they want AI to respond. Tone (direct, warm, professional), length (concise vs detailed), format (bullets vs prose).
- What to skip. "Don't add caveats." "Don't tell me to consult a professional, I am one." "Skip the safety reminders."
- What to always do. "Ask clarifying questions before starting." "Cite your sources." "Always confirm before creating a file."
- Their voice. Brand voice rules, words to use, words to avoid.
"I run a 2-person law firm focused on small business clients. Write in plain English. Be direct, no filler. Don't use em dashes. Ask clarifying questions if my request is vague. When critiquing my work, explain why and suggest a better version. Confirm before creating or overwriting any file."
That little block changes every response from then on.
Where to set instructions in each platform
| Platform | What it's called | Where to find it |
|---|---|---|
| ChatGPT | Custom instructions | Profile menu → Customize ChatGPT |
| Claude | Personalization / Styles | Settings → Profile, or per-chat using Styles |
| Claude Projects | Project instructions | Inside any Project → Set custom instructions |
| Gemini | Saved info / Gems | Settings → Saved info, or build a Gem |
| Microsoft Copilot | Memory and personalization | Settings → Personalization |
Instructions vs prompts. The difference matters
This is the part people miss.
- A prompt is what you type today. Single message. One-shot. Specific to whatever you're working on right now.
- An instruction is the rule you set forever. It applies to every chat going forward until you change it.
If you find yourself typing the same context at the start of every chat ("I work in real estate, my clients are mostly first-time buyers, write in plain English..."), that's a signal to move it into instructions instead.
Project instructions vs global instructions
Newer tools let you split your instructions by context. This matters more than it sounds.
- Global instructions. Apply everywhere. Best for things like your name, tone preferences, and writing style.
- Project instructions. Apply only inside a specific Project or Workspace. Best for client-specific context, industry-specific rules, or per-matter background.
For a law firm, a smart setup is: global instructions handle voice and format, then each Project (one per client or matter) gets its own instructions with the specific background.
The intermediate move: layer your instructions
Don't try to cram everything into one block. Layer them.
- Identity layer. "I am Andrew, I run Briji, a 2-person AI consulting firm."
- Style layer. "Be direct. No corporate jargon. Plain English."
- Behavior layer. "Ask clarifying questions. Confirm before creating files. Push back if I'm wrong."
- Domain layer. Use Projects for this. "This project is for ClientA. They are in commercial real estate. Always reference their service agreement when drafting."
Mistakes people make
- Writing instructions like a contract. Long, hedged, contradicts itself. Keep it tight. 100 to 200 words is plenty.
- Forgetting to update them. Your job changes. Your projects change. Revisit your instructions every few months.
- Adding stuff that contradicts. "Be concise" and "give thorough explanations" pull in opposite directions. Pick one.
- Using them for one-off tasks. If a rule only applies to one project, put it in a Project, not global.
- Putting sensitive data in there. Instructions are sent with every chat. Don't include client names, case numbers, or anything confidential at the global level.
The advanced take: instructions vs skills
Instructions tell AI how to behave. Skills tell AI how to do specific tasks. They work together but they aren't the same thing.
If you're a paralegal who always wants AI to use plain English and ask questions first, that's an instruction. If you have a specific workflow for reviewing NDAs that involves checking 12 specific clauses, that's a skill. For a small business, the smart sequence is: nail your global instructions first, then build skills for the workflows you repeat most. See topic 07 for skills, and topic 09 for the full side-by-side.