Master Prompt: How to Write One You Actually Reuse

A master prompt is one reusable block of instructions that tells an AI assistant who you are, what you're working on, and what a good answer looks like — so you stop re-explaining it at the start of every conversation.

It isn't an official term. No model provider ships a feature called "master prompt." What people mean by it maps onto things that do exist: a system prompt, a set of custom instructions, or a project-level context file. Knowing that matters, because the built-in versions are stored and applied automatically, and a master prompt you keep in a notes app is one you'll forget to paste.

The short version:

• What it is: a standing brief — your role, your context, your constraints, your format preferences, and examples of good output.
• What it's for: consistency. The same question asked on Monday and Thursday gets the same shape of answer.
• Where it should live: in your tool's own memory feature (custom instructions, a Project, a Gem, a context file), not in a document you paste from.
• The main mistake: writing a long one. A master prompt that tries to cover everything produces worse output than a short one aimed at a single recurring job.

What goes in a master prompt

Five parts, each doing a different job.

Don't skip the last row. Anthropic's prompting best practices describe examples as one of the most reliable ways to steer output format, tone, and structure, and recommend a few relevant, varied examples — 3–5 for best results — rather than a single one. Its guidance on system prompts makes a related point about role setting: a specific, concrete brief beats an abstract one. Telling a model to "write professionally" means almost nothing. Showing it paragraphs you were happy with gives it something concrete to match.

If you want the named techniques behind each of these slots, the free guides collected in prompt engineering PDFs cover them properly — this article is about the container, not the technique.

Keep it short, and keep it about one job

The instinct is to write a master prompt covering everything you do. Resist it. A brief that covers your writing, your analysis, your code, and your email produces mush, because most of it is irrelevant to whatever you asked and the model has to guess which parts apply.

Write one per recurring job. A master prompt for weekly reporting, another for customer emails, another for first-draft research. Each one short and specific. Three focused briefs beat one encyclopedia. AI prompts for business breaks down what that looks like per team, if you want a starting set rather than a blank page.

Where to actually store it

This is the part that decides whether you use it. A master prompt in a Google Doc is a master prompt you paste for two weeks and then stop pasting.

Every major assistant has somewhere to put standing instructions:

• Custom instructions apply to every conversation. Good for things that are always true — your language, your role, your general preferences — and bad for job-specific detail, because it applies everywhere.
• Claude Projects give each project its own knowledge and custom instructions, so the context applies inside that project and nowhere else. That's the right shape for the one-brief-per-job rule.
• Gemini Gems are the same idea: a customized version of the assistant with its own standing instructions you can return to.
• A context file, if you work in a coding tool. Claude Code reads a CLAUDE.md from your project and loads it into every session automatically — a master prompt that lives with the work instead of in your head.

The pattern is consistent: scoped beats global, and automatic beats pasted. Pick the narrowest container that covers the job, and let the tool load it.

Writing one in fifteen minutes

Don't compose it from scratch. Build it from work you've already done.

1. Pick one recurring task. The thing you ask for most often.
2. Find a few of your best past examples of it. Real ones, from your own files, and different enough from each other that the model doesn't just copy one.
3. Write four sentences of context — who you are, who it's for, what's always true, what the output should look like.
4. Paste the context and examples into a Project, Gem, or custom instructions. Not a notes app.
5. Ask the model what's missing. Give it your brief and ask what it still needs to do the job well. Its answer is usually a better edit than anything you'd have added yourself.
6. Use it for a week, then cut. Delete every line that never changed the output. Most first drafts lose a third.

Step six is where a good master prompt separates from a long one. Length feels like thoroughness and reads like noise.

What not to put in it

• Anything confidential you wouldn't want stored. Standing instructions persist. Check where they're saved before putting client names or personal data in one.
• Model-specific tricks. Formatting hacks tuned to one model version break on the next. OpenAI's prompt engineering guidance changes between model generations for exactly this reason. Keep your brief about your context, which doesn't change, rather than the model's quirks, which do.
• Instructions the tool already follows. "Be helpful" and "be accurate" are wasted lines.

When a master prompt isn't the answer

A master prompt fixes repeated context. It doesn't fix repeated work.

If the task has steps — pull the data, check it, format it, draft the summary — then no amount of briefing turns a chat window into a thing that does those steps. You'll still be driving it manually each time. That's the point where people either build something or give up and keep re-prompting.

That's the gap Taku is built for: an AI-native desktop workspace where you mirror an AI app or workflow someone already got working, run it on your own machine, and save it as something you run again — instead of re-briefing an assistant from scratch every week. The free app library is the quickest way to see whether the workflow you keep hand-driving already exists. Taku is in Beta, and the Mac app is available now.

Key points

• A master prompt is a reusable standing brief: context, role, constraints, format, and real examples of good output.
• It isn't an official feature. The built-in equivalents are custom instructions, Claude Projects, Gemini Gems, and context files like CLAUDE.md.
• Store it where the tool loads it automatically. A brief you have to paste is a brief you'll abandon.
• Write one per recurring job rather than one for everything, and cut it back after a week of real use.
• Examples are one of the most reliable ways to steer format, tone, and structure. Anthropic recommends a few relevant, varied ones rather than a single sample.

FAQ

What is a master prompt in AI?

A reusable set of instructions you apply to every conversation about a particular kind of work, covering your context, your constraints, and what good output looks like — so you don't re-explain it each time. It's an informal name for what tools implement as system prompts, custom instructions, or project instructions.

How is a master prompt different from a system prompt?

A system prompt is the technical mechanism: instructions supplied to the model separately from the user's message, usually through an API. A master prompt is the everyday version of the same idea — the standing brief you write, whether it ends up in a system prompt field, custom instructions, or a Project.

How long should an AI master prompt be?

Shorter than you expect. A focused brief of a few hundred words aimed at one recurring task outperforms a long document covering everything you do. Cut any line that hasn't changed an output.

Where should I save my master prompt?

In the assistant's own feature for standing context — custom instructions, a Claude Project, a Gemini Gem, or a project context file. Saving it in a separate document means remembering to paste it, which is why most master prompts stop getting used.

Can one master prompt work across ChatGPT, Claude, and Gemini?

The context part travels fine — your role, your audience, your constraints, your examples are model-independent. Formatting tricks tuned to a specific model don't travel, and often don't survive that model's next version either.