Can AI build my pregnancy food checker before lunch?

This is my actual next App Store idea: type a food, learn whether it's generally safe in pregnancy, with a source for every answer. Before spending three months on it, I gave Claude Code one prompt and a list of eight FDA, ACOG, and CDC pages it was allowed to use.

This is my ACTUAL next App Store idea. You type a food, and it tells you whether it's generally considered safe in pregnancy, with a source for every answer. Before I spend three months building it, I gave Claude Code one prompt. It came back in five minutes with ninety rules and passing tests. Then I clicked one Source link, and the page it cited NEVER mentions the food. Here's exactly what happened.

Before we start

• Not medical advice
• Rules come only from FDA, ACOG, and CDC pages
• Ask your doctor or midwife

Two things before we start. NOTHING in this video or this app is medical advice. The app is a lookup over public guidance from the F D A, the American College of Obstetricians and Gynecologists, and the C D C, and every rule links to the page it came from. If you are pregnant and unsure about a food, ask your doctor or midwife. That's also the whole point of today's build. In a health app, the source link is the PRODUCT.

Why this build

• The idea from my planning doc
• Cheapest possible validation: one morning
• The question: does AI invent rules or sources?

Some context. This idea has been sitting in my planning doc for MONTHS. A pregnancy food scanner, eventually with a barcode camera. The cheapest way to find out if it's worth three months is to let AI build the smallest version in one morning and see where it breaks. And for a health app, the thing I'm watching for isn't the layout. It's whether the AI invents a rule, or invents a SOURCE.

The prompt, word for word

Here's the prompt, and it's in the description. A single file web app. A search box with instant results. Rules in a table with a food, a category of safe, caution, or avoid, a one line reason, and a source U R L. At least sixty foods, using ONLY eight pages I listed from the F D A, A COG, and the C D C, and every row must carry the U R L it came from. A permanent disclaimer. If a food isn't in the table, say so and don't guess. And a test file that checks specific answers and runs in the foreground.

The build, replayed

THREE in the morning, prompt sent. Yes, still not lunch. This is the real session replayed. It wrote the app in one file with the rules table inside it, wrote the test file, installed a browser for the tests, and ran them. EIGHT turns, five minutes, ninety rules, eight tests, all passing. I had not typed anything.

First run, untouched

First run, untouched. Type tuna and you get FOUR answers at once. Bigeye tuna, avoid, high mercury. Albacore and yellowfin, caution, one serving a week. Canned light tuna, safe. Type brie and it splits pasteurized from unpasteurized, which is exactly the distinction that matters. Coffee, caution, under two hundred milligrams of caffeine a day. And pizza, which isn't in the table, gets an HONEST answer. Not in our list, we don't guess, ask your doctor. Calm, big type, works on a phone. So far, better than I expected.

Then I did what a user would do

• Pick a rule, click Source, read the page
• Three rules for alcohol, one source
• Twenty two rules, one Source link

Then I did the thing a real user would do. I picked a rule and clicked SOURCE. Alcohol, beer, and wine were all marked avoid, which is correct advice. All three cited the same A COG healthy eating page. And I noticed something else. Twenty two different rules, from raw sprouts to rare steak, all linked to ONE single page.

The bug: a source that never says it

Here's the alcohol source. This is the actual A COG page, and I searched it for the word alcohol. ZERO matches. The advice is right, and the citation is WRONG, and in a health app that is worse than no rule at all. Because the whole promise of the app is that you can check for yourself. I didn't tell the AI anything technical. I said, I clicked Source on alcohol and the page never mentions alcohol.

Twenty two rules, one menu page

And the second problem. Raw sprouts, Source. It lands on the F D A Safe Eats page, which is a MENU. Meat, dairy, produce, ready to eat foods. The actual guidance is one click deeper, on the sub pages. Twenty two rules were citing the menu. Real rules, real agency, wrong page. So the second half of my report was, make every Source go to the page that actually states the rule, and if a rule can't be backed by one of my eight pages, DELETE it rather than keep it.

The fix, replayed

This is where it got interesting. It didn't just edit links. It fetched all eight pages ITSELF to read what each one actually says. Two A COG pages blocked it, the same block I hit. Then it went through the table rule by rule. THIRTEEN rules removed, including alcohol, because none of the eight pages I allowed says a word about alcohol. Around twenty rules moved to the C D C pages that name them. Twenty new fish added from the F D A list to keep the count up. And three new tests. No rule may cite a menu page. Every rule must cite one of the eight allowed pages. The table must stay above sixty rules.

After the fix

After the fix. Type alcohol and the app says it's not in our list, ask your doctor. Which is the CORRECT behavior for a rule it can't source, even though the advice is common knowledge. Type sprouts, click Source, and it lands on the C D C page, on the paragraph that ACTUALLY says avoid raw sprouts. That's what the app was supposed to do from the start.

Eleven checks, and one more message

I reran the tests myself. ELEVEN for eleven. Then I checked its homework. Its summary said ninety six rules. The file had one hundred and four. And it had deleted THREE rules that the C D C page does support, Caesar dressing, cookie dough, and cream cheese, because it was being cautious. One more message put those three back with the right citation. Twenty minutes, three prompts, done.

Verdict

• Does the idea survive? Yes, as a lookup
• The App Store version needs a clinician reviewer
• Never trust a Source link you have not clicked

Verdict. Does the idea survive contact with a real build? YES, as a lookup over agency guidance, which is what it should be. The App Store version needs three things this one doesn't have. A clinician to review the table, a barcode camera, and an alcohol source, because the rule people ask about most is the one the app now refuses to answer. And the lesson for anyone building with AI. It will happily write a citation that looks perfect and points nowhere. You don't need to read code to catch that. You need to CLICK the link.

Scorecard

Scorecard. Prompt at three, working app at three oh five, fully sourced app at three twenty. TWENTY minutes, three prompts. The bug wasn't a wrong rule. Every rule I checked was correct advice. The bug was wrong CITATIONS, and in a health app that is the bug that matters. Its fix was the right instinct. Delete anything it couldn't source, even the alcohol rule, rather than keep a link that lies. Cost, nothing on top of the Claude plan I already pay for, and the app itself calls no A P I at all, so it costs nothing to run.

The prompt, word for word

Next: a Chrome extension that catches overpriced groceries

This post is the episode's script. The narration in the video is an AI clone of the host's voice; the apps were generated by Claude Code and are shown as recorded. Tools: Claude Code on a paid Claude subscription, ElevenLabs on a paid plan, Remotion (free for individuals). Not sponsored.