Working with AI: prompts to copy

These are the prompts from my Working with AI sessions, gathered in one place so you can come back to them. They work in Microsoft Copilot chat and in Copilot Cowork, the part of Copilot that takes on a whole piece of work for you.

Copy a prompt, then change the words in brackets to fit your work.

Before you paste. Read your organisation's AI policy first, and follow it. If you would not put it in an email to a broad audience, do not put it in AI.

Set it up once: teach it who you are

Set up your context and voice once, and every task starts already knowing you.

1. Build your context card

“Tell me what you already know about me from our past interactions.”

Start in Copilot chat. Fill in the gaps it misses.

2. Teach it your voice

“Look at the emails I have sent in the last six months. What rules do I consistently apply?”

Keep what it gets right. Correct what it gets wrong.

3. Make it standing

“Keep this as standing instructions that you use every time you write on my behalf.”

Paste your finished card and voice guide into Cowork.

Brief in five lines, then improve it

Missing lines become assumptions. Wrong assumptions become rework.

  • Task: What do you need done?
  • Audience: Who is it for, and what do they care about?
  • Purpose: What should it achieve?
  • Constraints: Length, tone, template, deadline and no-go areas.
  • Output: Word, PowerPoint, Excel or email, and where it is saved.

The one-line habit

“Before you take any action, improve this prompt and show it to me. Wait for my go-ahead.”

It fills the gaps it would otherwise guess. Correct it once, then say go.

Test a draft before anyone else sees it

Check in proportion to the risk. Never send anything you have not read.

Compliance lens

“Scrutinise this as if looking for legal or compliance faults. Warn me about wording issues or risks.”

Stakeholder preview

“Act like [name and title] and scrutinise this. What feedback am I likely to get?”

Devil's advocate

“What haven't I covered? What would you ask? Debate the points I've raised.”

Policy check

“Is this compliant with [policy or legislation]? What would you add or change?”

Ask for a council, not a cheerleader

AI tends to side with whoever is asking. For a real decision, put it to five advisers, each testing it from a different side.

  • The Contrarian: “What breaks first?” Only looks at what will fail, and the early warning sign.
  • First Principles: “Starting from zero, would I still choose this?” Tests every assumption: verified, plausible or untested.
  • The Expansionist: “What is the bigger version of this bet?” Finds the upside you are missing, and an Option C.
  • The Outsider: “Why is it done this way at all?” Asks the plain questions insiders stopped asking.
  • The Executor: “What will you do on Monday morning?” Turns it into Monday's first moves and a two-week test.

Use it for real decisions. Run it in Copilot chat.

The council, in short

Run a decision council for me, so I get challenge rather than agreement.

First, in one message, ask me what the decision is, the context, what a good result looks like in a year, my constraints, my deadline, which way I am leaning and what worries me most. Wait for my answers.

Then have five advisers answer separately: the Contrarian (what fails first), First Principles (which assumptions are untested), the Expansionist (the bigger opportunity, and an Option C), the Outsider (the plain questions a stranger would ask) and the Executor (Monday's first steps and a two-week test).

Next, have the advisers review each other's answers without knowing who wrote which.

Finish as the chair with one clear decision, your confidence, the biggest risk, the next steps and the signal that would tell me to change course. Do not flatter me, and do not change the decision when I push back unless I give you new information.

For the full prompt, with the interview questions, the blind review and the chair's report set out in detail, use Mariah Brunner's free guide, The LLM Council Prompt. The council builds on Andrej Karpathy's LLM Council (2025) and Ole Lehmann's five-adviser version for Claude (2026).

Feedback is how it learns your standard

AI looks for patterns. Correct it, tell it what changed, and save the lesson.

“Compared with your draft, here is the version I sent. Tell me what I changed, then save it as my preference.”

Meetings become manuals and answers

A recorded Teams walkthrough can become a step-by-step guide anyone can search. Try these on a meeting transcript.

“Map the end-to-end process flow we discussed, based on the transcript.”

“List the exact questions I asked, each followed immediately by the answer.”

Happy to take you or your team through these. Get in touch at helen@helenhovenga.com.