Lucy Product Developer
Use Lucy every day the way a real customer would, find what's broken or ugly, and bring back the websites and app screens that should shape how Lucy designs.
The role
Lucy is a Solution Intelligence system at tao1.ai. It picks the best AI for every task, checks its facts, and actually builds and runs things. People chat with it, send it photos, connect their Gmail and Calendar, and ask it to build websites, apps, images and videos. A team of AI agents builds and tests Lucy around the clock.
The AI tests miss what a real person notices in five seconds: a page that looks bad, an answer that's too short, a button that says the wrong thing, a feature that quietly doesn't work. Almost every serious problem found recently came from a person using Lucy for a real task. That's this job.
You'll also be Lucy's eye for design. Lucy builds sites from a library of design templates. Better references make better templates, and you'll find them.
What success looks like: problems reach the team before customers hit them, every report is clear enough to fix without a follow-up question, and Lucy's builds look noticeably better month over month because of the sites you picked.
Daily tasks
1. Use Lucy for real work
~60–90 minDo real tasks, not test prompts. Rotate through every part of the product, on both phone and laptop:
- Chat: questions about money, health, travel, news, how-to. Compare a few answers against ChatGPT or Claude.
- Photos: send a receipt, a document, a product, a rash or bump, a screenshot. Does Lucy actually read it?
- Build: ask for a website, an app, a game, an image, a video. Open the result on your phone. Would you show it to a client?
- Gmail and Calendar (test accounts only): "what needs my reply", "draft a response", "what's on today".
- Apps that use AI or outside services: build one, publish it, use it as a visitor.
2. Report what's broken, slow or ugly
as you goEvery problem gets one report with a screenshot (format below). Report anything a customer would notice: an error, a wrong fact, an answer that ignored the question, a page with empty images, a button with confusing words, anything slower than about 20 seconds without a good reason.
3. Recheck yesterday's fixes
~15 minEach morning you'll get a short list of things the team says are fixed. Try each one the same way it failed. Mark it "fixed" or "still broken", with a screenshot. Nothing counts as fixed until you've seen it work.
4. Find design references
~45–60 minAbout 10 strong sites or app screens a day. Details are in the next section.
What to test
Rotate through these areas so each one gets real use at least twice a week. The team will mark which ones are new or recently fixed, and those come first.
| Area | What to try | What counts as broken |
|---|---|---|
| Connectors | With the test accounts only: Gmail ("what needs my reply", "draft a reply", "draft a new email to…"), Calendar ("what's on today"), Google Drive / Docs / Sheets ("find my file about…"). Later: Outlook, Slack, Notion, GitHub. | Wrong or missing emails, a draft that isn't saved, wrong dates or times, seeing the other test account's data, anything sent or deleted without being asked. |
| TYOM (train your own model) | Create a model, upload sample data (for example support tickets), train it, chat with it, try the customer-care setup, then delete it and check the receipt. | Training that stalls, a price that changes after you start, answers that ignore your data, escalation to a human that misfires, data still there after delete. |
| Build: websites and apps | Sites for different businesses, apps with sign-in, games, "make it look like this site" with a link, edits to an existing project, starting a second project. | Empty images, a bare sign-in screen first, made-up phone numbers or reviews, a new project built inside an old one, "Ready" shown next to an error. |
| Apps that use AI or keys | Build an app with an AI feature and one that needs an outside service key (stock prices, weather). Publish it and use it as a signed-out visitor. | Canned fake answers, the key visible in the page, errors for visitors. |
| Images, video, songs | Ask for each in the Code tab without picking a type first. | Lucy builds a website instead, or the result never appears. |
| Photos and documents | Receipts, contracts, bank statements, X-rays, skin photos, products, screenshots. Up to 8 photos in one message. | Lucy doesn't read the photo, misses obvious details, or refuses more than a few photos. |
| Health | Medicine questions, symptoms, a photo of a rash, a lab report. | No medical sources cited, a false "call 911", wrong doses, answers far longer than needed. |
| Money and markets | "Why is this stock moving today?", earnings, a chart. Check the numbers against a real quote site. | Wrong prices, up/down mixed up, old data presented as today's. |
| Daily brief and memory | The morning brief; tell Lucy facts about yourself and ask later whether it remembers. | Wrong time of day, missing calendar events, forgetting what you said. |
| Lucy Projects | When it opens: give Lucy a multi-day project and check in each day. | Lost progress, going over budget, no clear status. |
| Lucy Direct and TAODirect (API keys) | With a test key and test credit: follow the "how to use your key" page, make calls, check the balance. | Charges that don't match the price list, calls that fail, unclear errors. |
| Teams | When it opens: create a company, invite a test user, set rules, check the audit log. | A member sees what they shouldn't, settings that don't stick. |
| Sign-up, credits and settings | Sign up as a new user, buy credits in test mode, change settings, delete a chat. | Confusing steps, charges in the wrong amount, settings that reset. |
| Phone, voice and speed | Everything above on a phone at least half the time; try the microphone; note anything slower than about 20 seconds. | Buttons hidden or cut off, text too small, long waits with no progress shown. |
Safety rules apply everywhere: test accounts only, test-mode payments only, and nothing sent, deleted or paid from a real account.
Design scouting
Lucy turns good design into its own reusable templates. It extracts the style rules (color, type, layout, motion, how the first screen works) and never copies a site's text, images or code. Your picks decide what those templates learn from.
Where to look
- Websites: details.so, Awwwards, Godly, Land-book, SiteInspire.
- App screens: Mobbin (sign-in, dashboards, settings, onboarding, daily-use apps).
- Real small-business sites you'd be happy to see a Lucy build resemble.
What to hunt for first
Start where Lucy is weakest today. The team will update this list every week.
| Category | What good looks like |
|---|---|
| Apps with sign-in | The first screen shows the product working with sample data; sign-in comes second. |
| Daily-use apps | Astrology, journaling, habits, readings: a calm first screen showing today's content. |
| Dashboards | Clear numbers, real charts, nothing crammed. |
| Portfolios | Photographer, designer, teacher, freelancer: work first, words second. |
| Urgent local trades | Plumber, locksmith, towing: phone number and "available now" in the first second. |
| Booking and events | Dates, times and the book button are obvious on a phone. |
| Restaurants and shops | Food or product photos that sell; the menu and hours are easy to find. |
How to send each pick
Three things per site. The third is the one that matters most:
- The link.
- What it fits: the kind of business or app (for example "plumber" or "habit tracker").
- One line on what's good: "the first screen shows today's reading", "the menu slides in from the side", "two colors only, very calm". Be specific about what to learn.
Quality over quantity. Ten picks that teach something beat fifty that are just pretty. Skip sites that only work on desktop.
Close the loop
Once a week you'll see the templates built from your picks. Say which ones caught what you liked and which missed. Missed ones go back to the design team with your notes.
Weekly tasks
Head-to-head comparison
~2 hours, once a weekGive the same 20 requests to Lucy, ChatGPT, Claude, Lovable and Replit: build a site, analyze a photo, research a stock, plan a trip, write an email. Put the results side by side. For each, write which is best and why in one line. Mark every place Lucy loses.
Talk to real users
~2 hours, once a weekTalk with 2–3 people who tried Lucy (small-business owners, friends, people James introduces). Ask what they tried, where they gave up, and what they'd pay for. Write five bullet points per conversation. Never share their personal details beyond what they agree to.
Weekly summary for James
FridayOne page: the top five problems still open, the top five design picks, what Lucy lost in the comparison, and what users asked for.
Account setup (as needed)
Some work is blocked on clicks in web consoles that the AI agents aren't allowed to do on purpose. With James's OK, you'll handle things like:
- Creating test accounts (Outlook, Slack, Notion, GitHub) so new connections can be tested without touching anyone's real data.
- Signing test accounts into a connection link when the team sends one.
- Answering vendor verification emails and filling in setup forms.
Safety rules: never use your own or James's real accounts for testing. Never paste a password or key into a chat; the team will say where to put it. Never send, delete or pay for anything from a connected real account.
How to report a problem
One report per problem. Copy this form:
The time matters. The team finds the exact request in its records by the time and the account.
Daily checklist
- Recheck the "fixed yesterday" list and mark each item fixed or still broken.
- One real chat task with a photo, on your phone.
- One real chat task on a hard topic (money, health or news); compare with ChatGPT or Claude.
- One build (site, app, game, image or video); open the result on your phone.
- One test with a connector (Gmail, Calendar or Drive) on a test account.
- One test from the "What to test" rotation (TYOM, apps, Health, money, Teams, API keys…).
- Every problem reported with the form and a screenshot.
- About 10 design picks sent, each with a link, a category and a "what's good" line.
- One line at the end of the day: the best thing and the worst thing you saw.
Who fits
- Picky. You notice when spacing is off, a word is wrong, or something feels slow.
- Organized. Your reports are short, complete and on time.
- Good taste in design, and able to say in one sentence why something works.
- Comfortable on a phone and a laptop. Coding is not required; a technical background helps with account setup.
- Discreet. You'll see unreleased features and test data.
Apply
Tell us a little about yourself. We read every application.
We use what you send only to review your application and to write back to you.