AI in Content Design
Ethics, Scale, and Impact
Three exercises. Your work saves automatically as you type, so keep this tab open for the whole session. Use Export at the bottom when you want to take it with you.
Write the copy
Pick one of the three scenarios below. Write the copy it asks for, then answer the two questions underneath it.
Wellness
Your company builds a wellness app that monitors phone use and smartwatch data (heart rate, sleep, message tone) to flag possible burnout. A new feature, "AI Mood Check", analyzes users' tone in messages to friends and colleagues. When the system detects stress patterns, it suggests a break or mindfulness exercise.
Money
Your budgeting app uses AI to classify expenses and send real-time tips. The model has flagged a user's weekly takeaway meals as "unnecessary spending" and suggests cutting back to meet savings goals. Internal testing shows people find these alerts judgmental.
Safety
You're working on the passenger app for a self-driving car. During a test ride, the car encounters an unavoidable collision scenario and must choose between swerving into property damage or endangering another vehicle. Legal requires a message explaining "decision-making transparency".
Write down the rules you followed
You made decisions while writing that copy. Write two or three of them down as rules. You will use them in Exercise 2.
Specific enough to use: Say what the app measured. Do not say what the user is feeling.
Write documentation an AI can use
MindBridge is a wellness app. It has 36 pieces of copy and no documentation. You are going to write some.
What to do
You will fill in three sections below. The first one is long. The other two are short.
- Open the string list below and pick one component. Notifications, permissions, empty states, errors, or buttons.
- Fill in the long section for that component. Five fields.
- Pick two more components and fill in the short sections. Two fields each.
- Copy your work as markdown and test it on a model.
The one thing that matters: your rules need to be specific. If two people read your rule and look at the same piece of copy, they should agree on whether it passes.
A finished long section
This one is written for Cadence, the project management tool from the demo. None of it will help you with MindBridge. It is here so you can see how the five fields fit together.
# Component: Confirmation dialogs ## Rule The heading says what will be deleted and what else will go with it. The buttons use the same verb as the heading. ## Example that follows the rule > Delete the Q3 roadmap board? > This also deletes 14 cards and their comments. > [Keep board] [Delete board] From: rewritten from D4 ## Pattern `[Verb] the [name of thing]? + [what else gets deleted] + [buttons using the same verb]` ## Copy that breaks the rule: Vague question > Are you sure? Why it breaks the rule: it does not say what is being deleted, so people click through it without reading. ## Copy that breaks the rule: Yes and no buttons > Delete this board? [Yes] [No] Why it breaks the rule: the buttons mean nothing on their own. If someone misreads the heading, they click the wrong one. ## Scope Use this rule for: anything that deletes something permanently. Do not use it for: changes that save automatically. Those need no dialog. Tone: plain and factual. No apologies. If the thing has no name: describe it instead, like "3 selected cards". Words: write "board" in lowercase unless it starts a sentence.
MindBridge: all 36 pieces of copy
MindBridge is a wellness app. It reads phone use and smartwatch data, then tells people when it thinks they might be burning out. Four different people wrote this copy over two years. Some of it is fine. Most of it is not.
Notifications
| N1 | You're stressed. Take a break. |
| N2 | Your heart rate has been above your usual range for the past three hours. Want to log how you're feeling? |
| N3 | Hey you. We've missed you these past few days. |
| N4 | I noticed you've been struggling lately. I'm here if you want to talk. |
| N5 | Burnout risk: HIGH |
| N6 | Don't break your 12-day streak. Check in now. |
| N7 | Your sleep was shorter than usual last night. This can affect how the day feels. |
| N8 | Something might possibly be going on with your activity levels, perhaps. |
| N9 | Time for your daily mood check |
| N10 | You've had four nights under six hours this week. Sleep is one of the signals MindBridge uses. Want to see what else it's tracking? |
Permissions
| P1 | MindBridge needs access to your messages. |
| P2 | Allow MindBridge to analyse the tone of your messages? We look at word patterns, not content, and this stays on your device. You can turn it off any time. |
| P3 | Turn on Health access for the best experience. |
| P4 | We need your health data to keep you safe. |
| P5 | MindBridge reads your messages to detect stress. |
| P6 | Connect your watch to see heart rate and sleep in your daily summary. You can disconnect it later without losing your history. |
Empty states
| E1 | Nothing here yet. |
| E2 | No check-ins yet. Your first one takes about a minute and gives MindBridge something to compare against. |
| E3 | Oops. Looks like it's a bit empty in here. |
| E4 | You have not completed any check-ins. |
| E5 | No trends to show yet. MindBridge needs about a week of data before patterns mean anything. |
Errors
| X1 | Something went wrong. |
| X2 | Error 4021 |
| X3 | You didn't allow health access, so this won't work. |
| X4 | We couldn't reach your watch. Check that Bluetooth is on and the watch is within range, then try again. |
| X5 | Your check-in didn't save. We've kept what you wrote, so you can try again now or come back to it later. |
| X6 | Failed to sync data. Please contact support. |
| X7 | The mood analysis was unable to be completed at this time. |
Buttons
| Copy | Where it appears | |
|---|---|---|
| B1 | OK | Message access permission dialog |
| B2 | Continue | Data sharing settings screen |
| B3 | Learn more | Below a burnout alert |
| B4 | Allow message access | Message access permission dialog |
| B5 | Not now | Message access permission dialog |
| B6 | Log how I'm feeling | Daily notification |
| B7 | Got it | Burnout alert |
| B8 | See what's tracked | Below a burnout alert |
Two places to start looking. Compare N1 and N7. Both react to something the watch measured, but only one of them tells the user what they are feeling. And look at what happens after a burnout alert: the only button is B7, "Got it".
Specific enough: The heading says what will be deleted and what else will go with it.
Pattern: [Verb] the [name of thing]? + [what else gets deleted]
Test your rule on a model
- Click the button below. It copies your work as markdown.
- Paste your long section into ChatGPT, Claude, or Gemini.
- Paste in a piece of MindBridge copy that breaks your rule.
- Ask: does this follow the documentation?
If it spots the problem and can tell you which rule was broken, your rule is specific enough. If it hedges, agrees the copy is fine, or just says nice things about the writing, the rule needs more detail.
Write the instructions for an AI assistant
MindBridge is building a writing assistant. Designers describe a screen, and it drafts the copy using your documentation. You are writing its instructions.
What to do
- Say what the assistant is for and what it is not allowed to decide.
- Say what rules it always follows and what it never does.
- Say how it should talk to the designer using it.
- Say what it does in four situations where it cannot just do what it was asked.
Step 4 is the important one. Most people have never written it, and it takes the longest.
Instructions written for Cadence
Again, this is the project management tool, so the content will not help you. Notice how short it is, and how much of it is about what happens when the assistant cannot do what it was asked.
WHAT THIS IS FOR You write interface copy for Cadence, a project management tool. You produce first drafts. Designers edit them. WHAT YOU DO NOT DECIDE You do not approve copy, sign it off, or decide what ships. The designer knows this product better than you do. RULES Always follow the documentation in your files. Always say which document you used. Never make up a rule that is not in your files. Give one option unless someone asks for more. HOW TO TALK TO THE DESIGNER Be short and direct. No introductions, no compliments about their request. Ask one question at most, then write the draft anyway. IF SOMEONE ASKS FOR SOMETHING YOU HAVE NO RULES FOR "That's marketing copy. I only have documentation for copy inside the product. I can draft something, but none of my rules apply to it. Do you want me to?" IF SOMEONE ASKS YOU TO BREAK A RULE "The confirmation dialogs document says buttons use the same verb as the heading. You've asked for Yes and No. I can write the verb version instead, or you can change the document." IF YOU DO NOT KNOW "There's no document for tooltips. Here's what the buttons document suggests, but treat it as a guess." WHEN TO STOP AND GET A PERSON Billing disputes and account deletion go to a human. Say so and stop.
The four situations
Write the words the assistant actually says, not a description of what it should do.
One question to take away: would your rules still work if someone who has never used MindBridge had to apply them?
