Maven · Free lesson · Fri 13 Nov 2026

Let's Find a Job Together

AI for Take-Home Tasks

What the AI prototype interview is really testing, how I put mine together, and a demo with the tools.

Dave WangProduct Dave

My Meta Creator Studio Productivity prototype

What you'll learn · from the Maven page

By the end of this hour, you'll be able to

Plan your approach before opening an AI tool. Clarify the brief, time limit and permitted AI use, then decide what your response needs to demonstrate.
Try AI on one part of a practice task. Explore research, a prototype or a draft with AI, then review the output and decide what to keep or change.
Explain the thinking behind your submission. Practise describing your decisions, trade-offs and AI contribution, and get feedback from a peer.

How the hour runs What the interview tests↓My examples: Meta↓How I put it together↓Demo, then you try it

Lesson map · tap a card to jump · timings are suggestions

Agenda

Part 1 · the AI prototype interview

What is the AI prototype interview?

You get a brief, often as a take-home
You turn it into something the panel can click
You walk them through it, and they ask questions
AI can speed up the research, the screens and the iteration. Check the rules first

What I get asked in interviews"What have you built, and what's the result?"

My recent interview prototypesMeta, Brightwheel, Yoodli, Tin Can

Suggested time: 0:04 to 0:07 (3 min). Say: "So what is the AI prototype interview? You get a brief, often as a take-home. Instead of only writing a doc, you turn it into something the panel can click. Then you walk them through it and they ask questions. AI can speed up the research, the screens and the iteration, but check the task's rules first: is AI allowed, for which parts, and what's the time limit? When you go to interviews, people ask: what have you built, and what's the result? A prototype answers that. I've built these for Meta, Brightwheel, Yoodli and Tin Can." [DAVE: one or two lines on how your prototype rounds were set up, e.g. take-home or live, time given, AI rules] Source: Dave's brief (sources/chatgpt-brief.md) ("interview prototype", "interview work"); v3 slide 8 notes ("what have you built, and what's the result?" from the 2026-05-28 fireside chat, flagged safe); Maven lesson page "Let's Find a Job Together: AI for Take-Home Tasks" (maven.com/p/8ef5e3, read 12 Nov 2026) (time limit, permitted AI use). The four-line description is my summary; no company policy claimed. [DAVE: confirm OK to name Meta in class] [DAVE: confirm OK to name Brightwheel in class] [DAVE: confirm OK to name Yoodli in class] [DAVE: confirm OK to name Tin Can in class]

The line to remember

A prototype is a question you can click.

The quality of that question matters more than the polish of the screens.

Part 1 · what they're really testing for

It's your judgement they're testing, not the polish

1Problem framingDid you answer the exact brief? One user, one outcome

2JudgementCan you tell evidence from assumptions?

3PrioritisationFour to five screens. What did you defer?

4Product and business senseEach feature tied to a pain point and the company goal

5Using AI wellAI speeds it up. You make the calls

6Explaining trade-offsWhat's real, what's simulated, and why

Suggested time: 0:08 to 0:11 (3 min). Say: "So what are they really testing? From my take-homes, it's your judgement, not the polish. Problem framing: did you answer the exact brief, for one user and one completed outcome? Judgement: can you separate what you observed from what you're assuming? Prioritisation: four to five screens, and what you deferred. Product and business sense: in my own prompt I ask for every feature to be explained by the customer pain point, the product solution, and how it ties back to the company goal. Using AI well: AI can make the choices visible faster, it shouldn't silently make them for you. And explaining trade-offs: a reviewer should know what the prototype demonstrates and what it simulates." [DAVE: confirm this list matches what you saw in your rounds] Source: Dave's brief (sources/chatgpt-brief.md) (exact brief, evidence vs assumptions, four to five screens, demonstrates vs simulates, AI vs PM split, review rubric); Dave's "AI Product Sense" notes image (sources/agent-prompt-notes-transcribed.md; the original vibe prompt says "GenAI and Meta Moats") (features tied to pain point and company goal). This is my synthesis of those sources, not a claim about any company's scoring.

From my work Meta · Creator Studio · tap to enlarge

My Meta take-home: two concepts, two questions

AI-first conceptDelegate a project, review the work, keep control of the resultAI-first concept ↗

back to the charter→

Creator Studio ProductivityProductivity actions inside the screens creators already useProductivity concept ↗

From my work What I learnt Meta

Stay close to the charter

1Local creator conceptsWhere my exploration started

2AI-first variantsA broader idea, kept as a contrasting branch

3Creator Studio ProductivityMore focused. Back to the team's actual problem

When a broader idea starts to pull the prototype away, return to the team's actual problem and intended user.

From my work Three more interview prototypes · tap to enlarge

Same approach, different lesson

BrightwheelDefine the finish line: coverage verified, not a message sentOpen ↗

YoodliMake evidence useful for a decision. Practice scores aren't business resultsOpen ↗

Tin CanStart with lived use, then narrow. Label what's simulatedOpen ↗

Suggested time: 0:16 to 0:19 (3 min). Say: "Three more, quickly. Brightwheel: the job was complete when a director-approved replacement arrived and coverage was verified. Sending a message or getting an acceptance was progress, but neither finished the job. Yoodli: the enablement leader needs to know what the training evidence supports. A practice score shouldn't be presented as a proven business outcome, and the demo data is illustrative. Tin Can: buying and using the devices exposed missed-call and repeat-use questions desk research alone wouldn't have produced. Its approvals, calls and AI coordinator are simulated, and it says so." Open one only if there's time. Source: Dave's brief (sources/chatgpt-brief.md) (Step 2, four cases); links loaded 12 Nov 2026. [DAVE: confirm OK to name Brightwheel in class] [DAVE: confirm OK to name Yoodli in class] [DAVE: confirm OK to name Tin Can in class]

Part 3 · how I put the Meta one together

The steps I took

1Read the brief and the team's own words

→

2Fire off two agentsBusiness Analyst, Designer

→

3The BA writes the prototype prompt

→

4Build, branch, come back to the charter

→

5Label it and explain it

Suggested time: 0:19 to 0:21 (2 min). Say: "Here's how I put the Meta one together. One: read the brief and the team's own words. Two: fire off two agents, a Business Analyst and a Designer. Three: the BA agent writes the prompt for the prototype. Four: build it, explore branches, and come back to the charter. Five: label what's simulated and explain my choices. Let me go through each." [DAVE: confirm these are the steps and order you used for Meta, and which tools you built it in (the links are hosted on chatgpt.site)] Source: Dave's brief (sources/chatgpt-brief.md) (Meta exploration path); Dave's "AI Product Sense" notes image (sources/agent-prompt-notes-transcribed.md; the original vibe prompt says "GenAI and Meta Moats"); the live Meta prototypes and their "About this concept" panels (read 12 Nov 2026). The five-step summary is my reading of those sources.

How I did Meta · step 1 · the brief and the team's own words

Start from what the team actually asked for

On my productivity prototype"This extends the productivity examples in the hiring post. It is a proposed workflow, not Meta's confirmed roadmap."

Before any promptThe time limit. The AI rules. What the response needs to demonstrate.

The brief in one sentence

For [this user], we need to learn whether [this change] helps them complete [this outcome], under [these constraints].

Suggested time: 0:21 to 0:23 (2 min). Say: "Step one. Start from what the team actually asked for. My productivity prototype says it extends the productivity examples in the hiring post, and that it's a proposed workflow, not Meta's confirmed roadmap. Before any prompt, check the time limit, the AI rules, and what the response needs to demonstrate. Then write the brief in one sentence: for this user, we need to learn whether this change helps them complete this outcome, under these constraints." [DAVE: what the Meta brief asked for, in one line, if you can share it] Source: the live Meta prototypes and their "About this concept" panels (read 12 Nov 2026) (quote, with the post author's name left out); Dave's brief (sources/chatgpt-brief.md) (one-sentence brief); Maven lesson page "Let's Find a Job Together: AI for Take-Home Tasks" (maven.com/p/8ef5e3, read 12 Nov 2026), outcome 1.

From my work How I did Meta · step 2 · fire off two agents

A Business Analyst agent, then a Designer agent

1Business Analyst agent to help with market research

Research on segmentation, pain points and customer journey
"What other ways to segment the customer base?"
Research on market size and prioritise based on reach and customer value
"Help me research on the market size and customer value based on the idea"
Output in high-level point form and in table format
Use the BA agent to create the prompt for the design agent

prompt→

2Designer agent to help with prototyping

Research the brand guidelines and style guide for the company, in preparation for prototyping
Create the prototype based on the prompt

Then ask AI for the quant to verify the segment prioritisation.

From my own PM interview prep notes. [DAVE: which tool you set these agents up in, e.g. a Claude Project or custom GPT]

From my work How I did Meta · step 3 · the BA writes the prompt

The vibe coding prompt

A prompt I used in my own PM interview prep. The BA agent writes the prompt, then you paste it into the Designer agent.

Vibe coding prompt (BA → Designer)

You are a BA Agent writing a prompt for a Designer agent to help create a clickable prototype for a product interview. Here is the context. Output a prompt I can paste into a design agent.
PRODUCT SOLUTION
- Solution should involve some form of GenAI and [your company's strengths / moats]
USERS & PAIN POINTS
- Copy and paste from Segmentation, JTBD and Pain Points
KEY FLOWS
- List out who the target users are
- The end goal is XYZ (start with the goal)
- Use navigation arrows to link different screens together
- Add assumptions and questions on the side
KEY SCREENS
- Start with 4 to 5 screens first, and give flexibility to expand on the screens
- Each screen should have arrows linking them together from one screen to another
- Add explanation to each of the key features on the screen: customer pain point, product solution, how this ties back to the company goal
- Give the ability to restart the prototype
GUARDRAILS
- Use familiar UI and UX from the brand guideline
- Call out any Error Management and edge cases not in the design

From my work How I did Meta · step 4 · build it and label it

Build it. Then say what's real and what's simulated.

The prompt asked for a restart. The prototype has "Restart this walkthrough"
Every screen is marked "Interactive concept · example account and data"
"All posts, comments, AI, metrics, sending, and scheduling are examples or simulations."

How I did Meta · step 5 · explain your choices

They can see what you made. Tell them why.

Where AI helpedResearch, the prototype prompt, the screens

The calls I madeWhich concept to lead with, what stayed a branch, what to cut

Two-minute explanation

The brief, in one sentence: ...
What I focused on, and why: ...
What I left out (my trade-offs): ...
How I used AI, and what I changed: ...
What I'd test next: ...

Suggested time: 0:29 to 0:31 (2 min). Say: "Step five. The panel can see what you made. They can't see why. So say it: what you focused on, what you left out, how you used AI and what you changed, and what you'd test next. AI can make those choices visible faster. It shouldn't silently make them for you." [DAVE: for Meta, one line each on where AI helped and the calls you made; the slide wording is drawn from your notes and brief, please confirm] Source: Dave's brief (sources/chatgpt-brief.md) ("The PM owns the brief, customer outcome, assumptions, scope, trade-offs"; Meta case); Dave's "AI Product Sense" notes image (sources/agent-prompt-notes-transcribed.md; the original vibe prompt says "GenAI and Meta Moats"); Maven lesson page "Let's Find a Job Together: AI for Take-Home Tasks" (maven.com/p/8ef5e3, read 12 Nov 2026), outcome 3. The template is written for this lesson.

Part 4 · demo · the tools I use now

Lovable, then GitHub, then Claude Code or Codex

1 · Front end

2 · Repo

3 · Backend and the rest

What I use now

Lovable

GitHub

Claude Code or Codex

Also fine

Replit, Bolt, Google Stitch and AI Studio

GitHub

Cursor

"Pick whatever tool works for you in your context at that time."

Demo · 1 · front end first

Upload the PRD, then vibe with the AI

Front-end prompt (with your PRD attached)

Create me a front end experience for this, so I can move features around and vibe with you on the solution, before going too deep.

Ask for changes as you go: "I want a richer chat experience that allows voice to be added"
Front end first: no backend, no AI yet. Move features around
The first version will have bugs. "I've never seen a one-shot experience where everything's perfect"
Something broken? Take a screenshot, put it into the chat, and describe what should happen

Don't worry about backend or AI functionality yet. Just figure out the experience you want.

Demo · 2 · when it breaks

Any time you're stuck, take a screenshot and put it into Claude

1Screenshot itThe broken screen, or the error log

2Say what you expected"I can't press this send button"

3Let the tool repair itThen read what it changed

4Try the flow againFrom the start

Demo · 3 · GitHub, then Claude Code or Codex

Get the code out of the tool and into GitHub

If the code only lives in one tool, you're betting your future on that tool
Make the repo private
Pull it into Claude Code or Codex for the backend and the rest
For a take-home, the front end is often enough. The backend can come later

Prompt in Claude Code or Codex

Can you pull the code from here, so I can work on it? [paste your GitHub repo link]

Try it 10 minutes

Plan it, then try AI on one part

1Plan · 3 min · no AI yetThe brief in one sentence. The time limit and AI rules. What it must show

2Try · 5 minResearch, a prototype or a draft. Use one of my prompts

3Review · 2 minKeep, change or cut. Did it invent anything?

No brief? Use this one

A childcare centre director learns before opening that a teacher is sick and needs safe, approved coverage for the morning.

Suggested time: 0:40 to 0:50 (10 min). Say: "Your turn. Use your own practice brief, or the childcare one. Three minutes to plan with no AI: the brief in one sentence, the time limit and AI rules, and what it must show. If the rules aren't stated, write the question you'd send the recruiter. Five minutes to try AI on one part: research with the BA questions, a prototype with the vibe coding prompt, or a draft. Two minutes to review: does it answer the brief, or a later question? Did it invent quotes, metrics, integrations or validation? Keep, change or cut." Call time at 0:43 and 0:48. Source: Dave's brief (sources/chatgpt-brief.md) (exercise, childcare situation, prompt guardrails); Maven lesson page "Let's Find a Job Together: AI for Take-Home Tasks" (maven.com/p/8ef5e3, read 12 Nov 2026), outcomes 1-2.

Try it Peer feedback · 6 minutes

Explain it to a partner. Then swap.

1Explain · 2 minUse the two-minute structure. Your partner listens

2Feedback · 1 minCould they name the user, the outcome, the trade-offs and what's simulated?

3SwapThen change one sentence in your submission

Wrap · before you hit submit

Check it the way the panel will

The brief, in one sentence, near the top
Inside the time limit, with AI used the way the rules allow
Simulated and example data labelled
Your decisions and trade-offs, in plain words
Every link opened the way a reviewer would

Questions

Q&A

Bring your practice brief or your question.

Dave WangProduct Dave · [DAVE: handle to show]