Skip to content
AI New Zealand
Our Playbook Find your Benchmark Executive AI Briefing AI Lunch & Learn AI Operating System Client Workspace Book Justin to speak AI Courses: Certify Your Team AI for schools: Schools & Students AI at Work Foundations course AI Leaders course Claude at Work course Microsoft Copilot course ChatGPT at Work course Gemini at Work course Putting Descript to work course Putting Midjourney to work course AI Roadshow Premium membership Toolkits Webinars Articles & Insights AI Academy Contact us

Free toolkit · The AI Toolkit

Meet your toughest customer before they walk out.

Your customers have already told you what they think, in reviews, complaint emails, support tickets and the quotes you lost. This toolkit has Microsoft Copilot or Claude read all of it, build an evidence-based twin of the customer you keep losing, and let you rehearse your pitch, your website and your quote on them first.

2 prompts, ready to fill in 5 steps to your own twin 120 pieces of practice feedback Copilot and Claude

What the mirror showed

The business thought it sold design. Its customers were buying certainty.

In the session we read a year of feedback for Tūī Lane Kitchens, an invented Hamilton kitchen company, then put its toughest customer through three trials: the pitch, the website and the quote. Every name and number in it is made up, and all of it is in the pack for you to try.

DESIGN

The business believed it sold design.

Customers already loved the design. It was praised even in the angry reviews, and it did not win or lose the job.

CERTAINTY

It lost customers on certainty.

When will it finish, what will it really cost, and who will tell me when something changes?

PRICE

Price was the decoy.

Loud, memorable and rare: 7 complaints in 120 pieces of feedback. Loud is not the same as common.

THE QUOTE

The better quote cost more, and won.

A week-by-week plan, honest allowances, one named contact and a fair deposit beat the lower number.

What is inside

Read the evidence, build the twin, then rehearse with it.

A four page guide, two prompts, a customer twin built for both platforms, and an invented business's year of feedback to practise on before your own data goes anywhere near AI.

The guide and the prompts.

The method on four pages, with the two prompts that do the work. Fill in the square brackets and paste.

  • Five steps to build a twin of your own
  • Stage 1: the prompt that reads your feedback
  • Stage 2: the template that turns it into a twin
  • Quick checks before you trust it
  • Three rehearsals and the privacy rules

Sam, the burned renovator.

The customer twin from the session, built twice, as a worked example to copy for your own.

  • The full Agent Builder instructions for Copilot
  • The same twin as a Claude skill, ready to upload
  • The WHY?, MIRROR OFF, MIRROR ON and RESET commands
  • Starter prompts and an opening line

The practice files.

Tūī Lane Kitchens is invented, and so is everyone in its files. Run the whole method on it once and you will know what a good result looks like.

  • 120 pieces of feedback, in Word and Excel
  • Reviews, emails, tickets, lost quotes and a survey
  • The sales script, for the pitch trial
  • The homepage copy, for the website trial
  • The original quote, for the quote trial

Build your own

Five steps from a folder of feedback to a twin you can trust.

Try it on the practice files first: the pack walks you through that in 20 minutes. Then do it for real, with the privacy rules beside you.

The five steps

  1. Gather the evidence. Six to twelve months of reviews, complaint emails, support tickets and notes on lost sales. Aim for at least 50 items, in one Word document, grouped by source, with dates.
  2. Strip it first. Names, phone numbers, email and street addresses, and order or account numbers come out before anything goes into AI. The twin needs patterns, not people.
  3. Use an approved work tool. Microsoft Copilot or Claude on your organisation's business plan, signed in with your work account. Never a personal AI account.
  4. Read the evidence. Run the Stage 1 prompt with your file attached. Then check it: read the quotes it cites, and delete any theme with no evidence behind it.
  5. Build the twin and rehearse. A Copilot agent with your feedback as its knowledge, or a Claude skill. Then check what it tells you with three real customers.

The privacy rules

  • Only use feedback customers gave you about your service, to improve that service.
  • Strip personal details before the data goes in.
  • Never build a twin of one real, named customer. Twins are composites.
  • Check it fits your own privacy policy and your obligations under the Privacy Act 2020.

The twin is a hypothesis, not the truth. It tells you what to test with real customers, not what to rebuild on its own say-so.

Three rehearsals

Let it reject your work before a customer does.

Paste in what you were about to send, a section at a time. The twin answers in character, pushes back on anything vague, and walks out where the evidence says a customer like them would.

THE PITCH

Your sales script.

Paste your script or your first-meeting talking points, one section at a time. Type WHY? when the twin pushes back.

THE WEBSITE

Your homepage.

Paste the copy and ask: “You've just found us on Google. Would you get in touch?” Then MIRROR OFF for a score.

THE QUOTE

Your quote or proposal.

Paste a recent one. If the twin rejects it, have AI rewrite it for that customer, then test the rewrite.

WHY?Steps out of character and lists the feedback behind its last reaction, with source and date.
MIRROR OFFScores what you showed it out of 10 against the customer's needs, with the top three fixes.
MIRROR ONSteps back into character, so the rehearsal carries on.
RESETStarts again, meeting your business for the first time.

Two platforms

Build it in the tool your business already pays for.

The prompts and the twin are the same in both. Only the set-up differs, and the pack covers each.

Microsoft CopilotClaude
Build it as An agent in Agent Builder, with the twin's instructions in the Instructions box. A skill: the same instructions in a SKILL.md file, in a folder you zip and upload.
The evidence Your feedback file as the agent's knowledge, uploaded or linked from OneDrive or SharePoint. Attached to the chat, or kept in a Project's files. The skill looks for it before it says a word.
Watch for Keep the feedback in Word, Excel or PDF. Agent Builder does not take Markdown files as knowledge. Skills need code execution turned on. On Team and Enterprise plans an owner may need to allow them first.

Whichever you pick, type WHY? often. A twin that cannot point to the feedback behind a reaction is guessing, and neither platform can fully stop it drawing on what it already knows.

Get the toolkit

The Customer in the Mirror Toolkit, free.

The four page guide with both prompts, Sam as a Copilot agent and a Claude skill, and the Tūī Lane Kitchens practice files: a year of feedback in Word and Excel, the sales script, the homepage and the quote. ZIP, about 550 KB.

We will email you occasionally about practical AI for New Zealand businesses. Unsubscribe any time.

Where this came from

Built live, on Friday the 13th.

This toolkit comes from The AI Toolkit, our weekly webinar. On Friday 13 November 2026 we built a fictional company's toughest customer in Copilot from a year of its feedback, then put the pitch, the website and the quote in front of them.

We build these live with members every Friday, on real workflows, in 45 minutes.

The AI Toolkit, weekly

A practical session every Friday, live in the AI Academy, building real workflows together. Premium members also get recordings of our best sessions.

AI Academy Premium →

Also free: the Project Activity Tracker, Sales Call Coach, On-Brand Documents and Organisational Brain toolkits, at newzealand.ai/toolkits.