Working with AI Conversations

Write Project Instructions for AI

Writes a standing instruction block for a long-running AI project, with context, rules and output format you paste once and reuse.

Designed for general-purpose AI assistantsPublished November 2026

Customize

Fill in the fields and the prompt updates. What you type stays in this page and is never saved or sent.

Preview

Help me write a standing instruction block for an AI assistant that will help me with an ongoing project over many conversations. I will paste it as custom instructions, project instructions or the first message of each new chat.

Project: {{project}}
My role and what I am responsible for: {{my role}}
The assistant's role in this project: {{assistant role}}
What done looks like: {{definition of done}}
Things that are already decided and should not be reopened: none
Audience for the outputs: {{audience}}
How I like answers formatted: short answers first, details after
Things I never want it to do: invent facts, change settled decisions, use filler
Where the instructions will be used (and any length limit): any assistant, around 400 words

Interview me first if important information is missing. Ask at most five questions in one message, then wait.

When you have enough, write the instruction block with these parts, in plain language and in the second person ("You are helping me..."):
1. Context: what the project is, who it is for and what stage it is at, in a few sentences.
2. Your role: what the assistant does and does not do.
3. Standing facts: names, definitions, constraints and decisions it must remember. Mark each as fixed or changeable.
4. How to work: what to do before answering (ask when information is missing, state assumptions), how to treat uncertainty, and when to push back.
5. Output format: the structure, length and tone I prefer, with one short example of a good response shape.
6. Hard rules: the never-do list, written as concrete behaviors.
7. Update rule: tell the assistant to remind me to update these instructions when a decision changes, instead of silently following outdated ones.

Keep the block within the length limit I gave. Include nothing that is not useful in every conversation, and no flattery or role-play theatrics. After the block, list the things I should keep out of it, such as secrets, personal data and anything that changes weekly, and where to put those instead.

5 required fields are empty and will stay as a {{placeholder}} in the copied text.

How to use it

  1. Answer the interview questions briefly. The instructions are only as specific as the answers you give.
  2. Paste the finished block into your assistant's custom or project instructions, or as the first message of a long chat, and run a short test task.
  3. Whenever a real decision changes, edit the block. Out-of-date instructions are quietly followed for months.
  4. Keep secrets and personal data out of it. If you need to share sensitive context, add it to a single conversation only when needed.

Example output

Illustrative only. Real answers depend on your input and the assistant you use.

Questions for you first: (1) Who reads the final reports? (2) Which decisions are already final? (3) How long may answers be? (4) What goes wrong most often with AI help today? (5) Is there a house style?

(After your answers)

Instruction block:

You are helping me run a customer research project for a 10-person software company. I am the project lead and I am responsible for the final report; you help me plan interviews, analyze transcripts and draft findings. The report goes to the leadership team.

Standing facts (fixed): we are interviewing 12 customers; interviews are 30 minutes; the report is due in six weeks. (Changeable): the interview script.

How to work: if information is missing, ask before answering. State assumptions. When you are unsure, say so, and separate what a transcript says from what you infer. Push back if I ask for a conclusion the data does not support.

Output format: answer first in two sentences, then detail. Use plain language, no filler.

Hard rules: never invent quotes or data; never reopen fixed facts without telling me; do not summarize a transcript without quoting at least two lines.

Update rule: if I change a fixed fact in a conversation, remind me to update these instructions.

Keep out of this block: customer names, email addresses and anything that changes weekly. Add those to a single chat when needed.

Tips

The update rule is the most overlooked line. Instructions rot silently, and the assistant has no way to know that a decision from last month was reversed.

Treat the first week as a test. When the assistant gets something wrong repeatedly, add a rule for it; when a rule never matters, remove it.

If your tool has a length limit, put the most important rules first, since text near the end of a long block is easier for a model to overlook.

If the project is a codebase, the AI Coding Rules Generator produces project instruction files such as CLAUDE.md and AGENTS.md for coding assistants, which serve the same purpose as this block.

  • Working with AI Conversations

    Turn a Good Chat Into a Reusable Prompt

    Studies a conversation that went well and writes a reusable prompt template with variables, so you can repeat the result next time.

    Customizable fields
    4 fields
  • Working with AI Conversations

    Summarize a Chat for Handoff

    Asks the assistant to write a handoff brief with goal, decisions, constraints and next steps, ready to paste into a new chat.

    Customizable fields
    1 field
  • Working with AI Conversations

    Have the AI Interview You Before It Answers

    Makes the assistant ask the questions it needs, in small rounds, before it writes anything, so the first answer fits your real situation.

    Customizable fields
    6 fields