Notion AI Instructions: Make Requests and Rules Clear

Notion AI reads the instructions you write at the moment you send them. There is no standing system prompt you configure once. Every request you make is its own instruction, which means the quality of the output depends directly on how clearly you write what you need each time.

This is different from AI tools that let you set a persistent persona or a saved instruction set. With Notion AI, the most effective approach is a template: a structure you copy into a block each time you start a specific kind of task. For project-level rules that need to be consistent — your tone guide, your naming conventions, the way you structure a summary — a tool like fenbs keeps the approved rules outside Notion, available to any connected assistant.

How Notion AI reads your instructions

When you invoke Notion AI on a block or a page, it reads the selected content and any instructions you type in the prompt. It does not read other pages in your workspace unless you have shared or embedded them in the current context. There is no user-level memory that carries rules across pages or sessions by default.

This means every request starts fresh. If you want Notion AI to write in a particular style or follow a particular rule, include that rule in the prompt or in the content the AI can see. Assuming it remembers last time leads to inconsistent results.

Write specific, structured requests

The most common reason Notion AI produces the wrong output is a request that describes the goal without describing the output. "Improve this paragraph" could mean shorten it, make it more formal, add a transition, or restructure it. Any of those is a valid improvement. Say which one you want.

A more specific request: "Rewrite this paragraph to be no longer than three sentences. Keep the main point. Remove examples." That has a clear constraint (three sentences), a clear preservation rule (main point), and a clear removal rule (examples). The AI has less room to guess.

When the output format matters, describe it before asking for the content. "Give me a three-column table with columns for Name, Status, and Owner. Populate it from the project notes below." The structure comes first, then the content source.

Use templates for repeatable tasks

For tasks you run on every meeting, every brief, or every handoff, write a prompt template and save it as a Notion template or a text snippet you can paste. The template should contain the instructions in full, not just a heading. Something like:

Summarize the meeting notes below.
Format: three sections — Decisions made, Action items, Open questions.
Each section is a bulleted list. No prose.
Do not include anything not in the notes.

With a template like this, anyone on the team can get a consistent summary without knowing how to prompt well. The quality is in the template, not in the person running it.

Set project-level context

If Notion AI needs to know something true about the project to do a good job, put that information in the prompt or in the page it is reading. Say your company name, your audience, your product category, or the constraint you are working within. "This is a summary for a client who is not technical. Avoid jargon" tells the AI something it cannot infer from the page alone.

For rules that need to stay consistent across your whole project — your brand voice, a list of terms to use or avoid, decisions already made — record them somewhere permanent. A block at the top of a shared Notion page is one option. An external tool connected to your AI assistant is another.

Handle inconsistent results

If Notion AI gives different results for similar inputs, the most likely cause is an ambiguous instruction. Find the word or phrase that could be interpreted more than one way and replace it with something specific. If you wrote "professional," decide whether you mean formal, precise, jargon-free or something else and use that word instead.

If you have a result you liked, keep it. Copy the instruction that produced it into your template library and note what made that output work. It is much easier to refine a prompt that worked once than to restart from a vague instruction. Two or three tested templates for your most common Notion AI tasks will save more time than any amount of prompting theory.

How to do this with fenbs

1. Create a free account and add a project for the work you are doing in Notion.
2. Record your standing project rules in Decisions and rules: tone, format requirements, decisions already made.
3. Add your best prompt templates as Skills in fenbs so any connected AI assistant can list and read them.
4. Connect Claude, ChatGPT or another supported assistant to fenbs over Integrations, and ask it to read the project context before working in Notion.
5. When a project rule changes, update the Decisions and rules entry so the change reaches every tool automatically.

For a broader look at Notion AI and how it compares to agent-based approaches, see Notion MCP vs. Notion AI. For writing general AI agent instructions, see AI agent instructions.