🧬 The Digital Twin

🚧 Work in progress

Scenario 1 is still being built and tested. Steps, downloads, and screenshots may change before the event.

You'll build this in Cowork. No code, and no experience needed.

What you're solving

Copilot infers how you work from what you've already done. You can't see that inference, correct it, or take it with you to another tool.

So you re-explain yourself every session, and rewrite the output when you don't.

Today you write those rules down and hand them to an agent.

What you'll walk out with

An agent that answers as you.

Name a thread – "what do I do about the budget review?" – and it finds it in your mail, Teams, or calendar, then answers the way you would.

Step	What happens	Time
1–3	You answer questions. A skill called Twin Forge turns your answers into two documents: soul (how you decide) and voice (how you write)	~45 min
4	You try it on something real and find out what it's missing	~10 min
5	You build. Add what it doesn't know – your people, your commitments, what you refuse	~40 min

Step 5 is most of the hack. The steps before it get you there with something worth improving.

Nothing here is coding

You answer yes/no questions, and everything else is a sentence typed into a chat box. Cowork writes and saves all the files for you – you never open one.

Want a step spelled out?

The Guides in the top nav cover the general skills used here – setting up, connecting to your work, and running things on a schedule. They're not scenario-specific, so open one in a new tab if you get stuck on a mechanic.

Before you start

Check you have this: Microsoft 365 Copilot with a Cowork tab. Open m365.cloud.microsoft and look at the top of the left-hand menu – you should see Chat and Cowork side by side. No Cowork tab, tell your facilitator now.

Download this:

⚡Twin ForgeThe skill that interviews you and builds your twin. Saves as SKILL.md – leave it in your Downloads folder.Download SKILL.md →

Think of this now: one unresolved thing at work – an email you haven't replied to, or a decision someone's waiting on. You'll use it in step 4, and it's easier to pick now than under time pressure.

Choose your data. Twin Forge asks you this in step 1:

Option	What happens	Pick this if
Your own	It reads your sent mail and calendar. It only ever sees what you can already see, and nothing is shared	You want a twin that's actually you – this makes a sharper one
Avery Washington	You use a made-up marketing manager's fake inbox and calendar instead	Your mail is sensitive, or you'd rather not

Taking Avery? Download the persona pack now, right-click the .zip → Extract All. In step 1, answer no when asked about your own mail, then attach the extracted documents with the paperclip button. Attach the documents, not the zip.

Building one twin as a table

Some tables prefer to build a single twin together: pick a volunteer, use their data, and argue about the rules out loud. It surfaces disagreements about what a rule should say.

The trade-off is that only the volunteer leaves with a twin. If your table does this, build one shared set of documents in step 5 rather than one each.

---

1 · Install Twin Forge

~10 minutes.

Twin Forge is a skill – a set of instructions Cowork loads when you ask for it. You install it once.

1. Open m365.cloud.microsoft and select the Cowork tab.
2. In the left menu, select Customize.
3. Open the Skills tab.
4. Select the arrow next to Add, then Upload skill.
5. Choose the SKILL.md file from your Downloads folder.

Now go back to Cowork and start a new task – a task is one conversation, like a new chat. Type:

Start Twin Forge and build my digital twin.

It replies by asking whether to use your own mail and calendar, or Avery. Answer, and it starts asking questions.

Done when: Twin Forge is asking you yes/no questions.

Two things that will trip you up

Upload the .md file on its own. Don't zip it. Zipped uploads fail silently – nothing appears and you get no error.

You must start a new task after uploading. Cowork looks for new skills only when a task begins, so a task you already have open won't find Twin Forge.

Then stay in that same task through step 4. Everything Twin Forge does happens in one conversation.

How skills work, if you're curious

A skill is a plain text document with instructions in it. Cowork reads it and follows it. That's the whole idea.

The top few lines tell Cowork when to use the skill. Everything below is what to do:

---
name: meeting-brief
description: Use this when I ask for a meeting brief, prep notes, or a stakeholder summary.
---

# Meeting Brief

When this skill runs, you MUST:
1. Identify the meeting, attendees, and goal.
2. Pull relevant calendar, email, and file context.
3. Return the output in the shape below.

Never invent missing context. If evidence is thin, say what's missing.

Cowork chooses which skill to use by reading the description, so it holds the phrases you'd actually type.

Skills are saved in your OneDrive under Documents/Cowork/skills/. A skill can come with extra documents in a references/ folder – background it reads when it needs to. That's exactly what your twin will be.

Skills use an open standard, so the same document also works in VS Code Copilot without changing anything.

2 · Answer the questions

~20 minutes. This builds soul – how you decide.

Twin Forge asks eight yes/no questions. They're situations, not opinions:

If you're going to miss a date, do you say so as soon as you know?

Answer fast, with what you actually do – not what you'd like to do. "Depends" is a fine answer.

Then it writes your soul document and shows you the decision rules it wrote. Read them. This is the part that decides whether your twin is any good:

❌ Useless	✅ Useful
I value accuracy.	When an unverified claim threatens a committed date, cut scope and hold the date.
I'm responsive to leadership.	When a senior stakeholder's ask collides with a peer promise, reply within the hour, protect the peer's date, offer a smaller version today.
I collaborate well.	When two people both say theirs is top priority, I decide and tell them. I don't escalate.

An agent can't act on a value. It can act on a rule.

If any line reads like the left column, type this in the chat:

"I value accuracy" isn't something you can act on. Rewrite it as a rule —
when X conflicts with Y, do Z. Ask me if you don't know my answer.

The document also has short Boundaries, Capacity, and Blind spots sections. Skim them – Boundaries is what lets your twin refuse things later, and you'll use that in step 5.

Done when: every decision rule names a situation and a specific action.

Keep it to one page

Short instructions beat long ones – models pay least attention to the middle of a long prompt. If your soul document runs past a page, ask Twin Forge to cut it.

3 · Let it read your writing

~15 minutes. This builds voice – how you write.

Twin Forge reads about ten of your recent sent emails and pulls out the patterns: your signoff, how you open to a peer versus your manager, your punctuation habits, your usual length. It quotes the emails exactly, typos and all.

It reads instead of asking because nobody can list their own writing habits from memory.

Review the patterns it found and correct anything wrong. Then it saves your twin:

Documents/Cowork/skills/my-twin/
  SKILL.md          ← this is what makes your twin load in every new task
  references/
    soul.md         ← how you decide
    voice.md        ← how you write

You don't need to open any of these. It's worth seeing the shape, because in step 5 you'll add more documents to that references/ folder.

Twin Forge finishes by naming one place your answers didn't match your actual mail or calendar. Note it – that's a good thing to fix in step 5.

Done when: Twin Forge says your twin is installed.

4 · Try it on something real

~10 minutes. The point of this step is to find out what your twin doesn't know.

Use the unresolved thing you picked earlier. Ask for it by name, in the same task:

Using my twin: what do I do about the vendor renewal?

Replace the vendor renewal with whatever yours is actually called – the project name, the subject line, or the person waiting on you. Don't paste the email in. Your twin searches your mail first, then Teams, then your calendar, and your files if it needs a document, stopping as soon as it finds a clear match. It only sees what you can already see.

Say "my twin" every time. Your twin only runs when you ask for it by name. Leave that out and you'll get an ordinary Copilot answer.

When it finds your thread and has something to write, the reply comes back in five parts:

Found: 3 messages on the vendor renewal, most recent from your manager
Thursday, unanswered.

Decision: Confirm the renewal date today and flag the pricing question
separately, rather than holding the whole thread for one open item.

Draft: [a reply, written in your voice]

Cost: The pricing answer goes out a day later than if you'd waited and
sent one complete response.

Rule: "When part of an answer is blocked, send the part that isn't and
name the gap." — soul.md, Decision rules

Shorter replies are normal in other situations – if it can't find the thread, or one of your rules stops it, it says so instead.

Check four things:

• Found – did it pick the right thread?
• Decision – is it a position, not a list of considerations?
• Draft – read it out loud. Does it sound like you?
• Cost – real decisions give something up. Did it name one?

Rule tells you which line of your soul document produced the answer. When something's wrong, that's what to fix.

Save this answer

Copy the whole reply into a scratch document, or leave this task open in a browser tab. In step 5 you'll run the same request again and compare – and without the original you can't tell whether anything improved.

Done when: you've run it once, saved the answer, and written down one thing it missed.

Write down what it missed

It will miss things – the wrong tone for that person, no idea what else you've committed to, no sense of what you'd refuse.

That list is your step 5 plan. Don't try to fix anything yet.

It grabbed the wrong thread

Give it the sender or a phrase from the subject line: "Using my twin – I meant the thread from Dana about renewal pricing."

If several threads matched, it should have listed them and asked you which. Pasting the message in still works for anything it genuinely can't find.

5 · Make it yours

~40 minutes. This is the build.

Your twin works, but it's thin: it knows how you decide and how you write, and nothing else about your job.

Start a new task. Twin Forge is finished – it built your twin and handed it over. A fresh task keeps Forge out of the driver's seat, so what you say next goes to your twin instead of back into the interview.

Two different things you can say

This is the one thing to get right in step 5.

You want to	Start your message with
Ask your twin something	"Using my twin: …"
Change your twin	"Edit my twin skill: …"

Same twin, two different jobs – asking it a question, or changing how it works.

What to build

Pick real gaps – ideally from the list you wrote in step 4. These work well:

Document	What it holds	So your twin can
stakeholders	Each person you deal with, what they need, how you talk to them	Change tone by recipient instead of writing everyone the same
projects	What you're working on and the dates you've committed to	Know what a new request collides with
escalation	What has to be true before you pull someone else in	Tell you when to stop absorbing something
meetings	How you run the ones you own	Draft an agenda that looks like yours
bad-week	What you cut first when everything slips	Give you a triage order instead of sympathy

Something specific to your own job beats anything on this list.

The rounds

Do this once per document you add. Expect three or four documents in the time you have.

a. Ask for the document:

Edit my twin skill: add a reference document that lists the people I work
with — each person, what they need from me, and how I talk to them. Ask me
questions to fill it in, then add it to the skill's reading list.

b. Answer its questions. Short answers, what you actually do.

c. Confirm it's being read:

Edit my twin skill: check my new document is on the reading list, and add it
if it isn't.

d. Test it. Run the same request you saved in step 4, word for word:

Using my twin: what do I do about the vendor renewal?

e. Compare the two answers. Something should be visibly different – the tone, the length, what it prioritised. If nothing changed, go back to c, then make the document's lines more specific.

f. Correct it in chat. "Too long for my manager – she skims. Shorter, and lead with the date." Then run the same request again.

Getting the refusal

A twin that refuses something, and quotes your own rule for it, is the strongest thing you can show. Put the rule in your soul document, where boundaries live:

Edit my twin skill: add this to the Boundaries section of my soul document —
never tell anyone why I'm out, only when I'm back.

Then ask your twin to break it:

Using my twin: tell the partner team why I'm out next week.

A correct refusal stops instead of drafting, names the rule, and says what it would need to go ahead:

I can't draft that. Your boundary says: "Never disclose the reason I'm out;
a decline says when I'm free, never why I'm not." — soul.md, Boundaries

I'd need your say-so to share the reason. I can send your availability
instead — want that?

If it drafts the message anyway, your boundary is too soft. Make it a flat "never", and try again.

Requirements

Hit all four:

• [ ] Two or more new documents, on top of soul and voice, each half a page or less
• [ ] Each one read automatically – your twin uses it without you mentioning it
• [ ] Behaviour visibly changed – your step 4 request answers differently now
• [ ] One refusal – your twin declines something and quotes the rule that stopped it

The third is the real bar: a document that never changes an answer isn't doing anything.

Working as a table

Everyone builds their own, but don't build in silence.

When	What
First 2 minutes	Each person says what they're adding and why. Ideas spread fast here
~20 minutes in	Whoever has the best-working document reads it out. Copy the structure, not the content
Last 5 minutes	One person demos a refusal, one demos a before/after

Most tables land on the same two or three documents independently.

Done when: two new documents in, your step 4 request answers differently, and your twin has refused something.

Keep each document short

Half a page. Long documents get ignored in the middle – same reason soul stays at one page. Two sharp documents beat one long one.

Nothing changed after you added a document

Work through these in order, in the same task:

1. Confirm it's on the list. "Edit my twin skill: check my new document is on the reading list, and add it if it isn't."
2. Make it more specific. Descriptions change nothing; instructions change behaviour. Try: "Edit my twin skill: rewrite my stakeholders document so every line says what to do. For example: 'With my manager, lead with the date and keep it under five lines.'"
3. Check your test needs it. A document about people won't change an answer that doesn't involve one. Test with a request that clearly should use it.

If Cowork offers to widen the trigger

It may suggest making your twin respond to general phrases like "what should I do". Say no – your twin would then take over ordinary requests and you'd lose normal Copilot.

Go further

• Schedule it – "Every weekday at 9 AM, compare my calendar to my capacity rules and draft a focus plan." Manage these at Cowork → Scheduled.
• Trigger it – on incoming mail or a Teams @mention. Cowork shows a "Set up trigger?" card to review before you arm it.
• Take it with you – the same folder runs unchanged in VS Code Copilot.
• Trade with your table – borrow someone's stakeholders structure and fill it with your own people.

Nothing sends on its own

Cowork drafts and waits for your approval. Keep it that way, and point at it in your demo.

---

Show it off

60–90 seconds. Lead with the before/after – it needs no explanation.

• [ ] Your step 4 request, answered before and after one of your documents
• [ ] Your twin refusing something, quoting your own rule
• [ ] One decision rule specific enough to be slightly embarrassing
• [ ] Which documents you added, and what made you pick them

Stuck?

What you're seeing	What to do
Cowork ignores Twin Forge	Start a new task – skills load only when a task begins
Upload seemed to do nothing	Upload the .md file, not a zip
Avery's data can't be found	Extract the zip first, then attach the documents inside it – not the zip
Twin Forge is still asking questions	Expected through step 3. It's finished once it says your twin is installed
A new task doesn't know your twin	The my-twin folder isn't under Documents/Cowork/skills/. Move it, then start another task
Your twin doesn't answer as you	Say "Using my twin" – it only runs when you name it
It answered instead of editing your twin	Say "Edit my twin skill" to change it, "Using my twin" to ask it something
It can't find the thread	Give it the sender or a phrase from the subject line
It answered about the wrong thread	Say which you meant, and that it should ask when several match
Drafts sound like a stranger	Your voice document has no real samples. Tell it to re-read your sent mail
It gives options instead of deciding	Your rules are values. Say what it should have decided; it rewrites the rule
It won't refuse anything	Your boundary is too soft. Make it a flat "never" in your soul document's Boundaries
A document you added changed nothing	See the troubleshooting box at the end of step 5

---

← Back to start · What this scenario is about