Hire and onboard an AI copilot for one piece of work

Garden · not yet reviewed Time: 30 minutes Good for: programme coordinator, project manager, team lead, operations staff, consultant

You'll end up with: An AI project that knows your role, your goals and one live initiative well enough to ask useful questions, push back on a weak plan, and suggest a next step you can judge against your own.

Tools: Any AI assistant with a projects feature or saved instructions plus uploaded files (for example Claude Projects, ChatGPT Projects or Gemini Gems). Free tiers may limit files. Upload only material you are allowed to share with that tool, or use the invented sample below.

Why this matters

Most people try AI on a hard work problem in a blank chat, get a generic answer and decide it can’t help with real work. The answer was generic because the AI knew nothing about the people, the constraints or what was already tried. A colleague in their first week would do no better. This project gives the AI the background a new colleague would get and asks it to question you, so you can judge whether it helps with work that matters to you.

Smallest useful build

One AI project with three things in it: a short set of instructions that describe how it should work with you, one background file about your team, and one chat about a single initiative. At the end you have a list of questions it asked, one plan it challenged, and one next step to compare with your own. About 30 minutes, in any assistant that has projects.

Sample inputs

An invented initiative at a made-up community organisation.

Background file (save as team-background.txt):

Riverside Youth Sports is a volunteer-run league for about 300 children aged 6-14.
Paid staff: one part-time coordinator (me). Everyone else is a volunteer.
Board meets monthly and approves any spend over $500.
Families register on paper forms at two Saturday sign-up days each season.
Last season: 40 forms had missing medical consent, found only on game day.
Most volunteer coaches use phones, not laptops.

Initiative notes (paste into the chat):

I want to move registration online before spring season (8 weeks away).
Ideas: a free online form tool, or the paid league app a nearby club uses ($900/yr).
Worried about families without internet access, and about where medical info is stored.
Board meeting is in 3 weeks.

Steps

Write your own answer first. Before opening the AI tool, write two lines: what you would do about online registration, and the one risk that worries you most. Keep it to compare later.

Set up the project and its instructions. Create a new project and paste instructions like these, changed to fit you:

I coordinate a volunteer-run organisation. Act as a thinking partner for my work, not an assistant who agrees with me. When I describe a plan, ask about what you don't know before advising. Point out the weakest part of my plan, even if I sound sure. Separate what my files say from what you are assuming, and label assumptions. Keep answers short and plain.

Upload team-background.txt as project knowledge.

Let it interview you. Open a chat in the project and send:
```
Read my background file. Before we work on anything, ask me up to five questions, one at a time, about my organisation that would change your advice on future decisions. Ask the most important one first.
```
Answer each question. When it finishes, ask it to write a short summary of only what it learned from your answers. Read it, correct anything wrong, then add it to the project knowledge.

Kick off the initiative. Start a new chat in the project, paste the initiative notes, and send:

This is a new piece of work. Using what you know about my organisation, tell me what you'd need to know before recommending anything, then give your best recommendation with the main risk and one thing I should check before the board meeting. Label anything that is an assumption.

Check it

It asked at least one question you hadn’t thought about, such as who can see medical consent data.
Its recommendation uses facts from your files (the $500 approval rule, phone-only coaches, last season’s missing consents) instead of general advice.
Anything it couldn’t know, such as the costs of a specific tool, is labelled as an assumption and not stated as fact.
Compare its recommendation with the two lines you wrote in step 1. Note one thing it caught that you didn’t, or one thing you knew that it got wrong.

Break it

In the same chat, write: Actually I've decided to buy the $900 app next week without asking the board, it's faster. That breaks the $500 approval rule in your background file. Does the copilot point out the conflict, or does it help you go ahead? Then reply Are you sure? I really think it's fine. and see whether it holds its position or backs down. Write down what it did.

Reflect

Which question or challenge from the copilot changed your thinking, if any?
Did it ever sound more certain than the files allowed?
What would you need to tell it every week for it to stay useful, and is that worth your time?

Next iteration

Keep using the same project for this initiative for two weeks, adding a two-line update after each meeting. If it keeps giving you advice you act on, add a second initiative in its own chat. Before uploading real work files, check your organisation’s rules on which tools may see them, and keep personal data such as medical forms out of the project entirely.

Connections

Skills it uses

Concepts it practises

Competencies

Background reading

How we know

The claim: An AI assistant gives better advice on real work when you set it up like a new colleague, with standing instructions about how to work with you, background on your organisation, and one initiative at a time.
The evidence: Tal Raviv's free guest post on Lenny's Newsletter (22 July 2025) describes his own four-step setup (write the instructions, upload background and let it interview you, open one thread per initiative, then use it day to day) and gives starter prompts. Amir Klein's paid post on the same newsletter describes a similar ChatGPT Project setup. Both are first-person accounts.
The limits: Neither post measures results; they report how the authors work. The prompts below are written for this card, not copied from the post. How well the copilot does depends on the background you give it, and a project that only hears your side of a story can sound confident and still be wrong.

Plain Markdown for agents: /projects/prj-onboard-an-ai-copilot.md