Cheap to build, hard to stop: when AI builds faster than you can learn

Fold started as decluttering. I wanted one master copy of the recipes I actually cook from my cookbooks, readable on an iPad and printable, in a form that suits a household where not everyone eats the same. Nothing I could buy did that, so I built it. It has become something more specific: a vegetarian kitchen that knows who's vegan, shows how each recipe could be made vegan where possible, where to split it into two pans, and how much each person needs. It's deliberately not "for vegans"; it's for the mixed table.

How it's built

Two Claude sessions share one folder. A chat session handles product, design, specs and research, and writes each piece of work as an UPDATE file straight into the repo (84 by 1 November). Claude Code does the engineering, branches, tests and merges, and sends REQUEST files back when it needs a decision. I coordinate and own the decisions. Small fixes take a fast lane, where CI merges them when checks pass, "so fixes ship even when Emily is out of credits"; features take a gated lane, and anything that changes stored data waits for me.

At its peak, 181 commits reached the main branch in a single day (28 October). No person could review that line by line, which is why Fold has a review panel.

Never loss-making

The rule: no free action triggers inference, and every AI cost is prepaid by the household that causes it, under a hard global ceiling. I wrote it down as a rule on 29 October, but the guards came first: a per-household allowance from 17 October (now 60 recipe reads to start, then 20 a calendar month), and from 24 October a global daily cap on Claude calls, taken atomically so parallel requests can't both slip past. If the cap can't be counted, nothing is called. The family and the friends' beta are subsidised by me; the rule applies from the first public release. October's bills came to about £72, most of it one-off setup (the ledger). The honest caveat: Fold's own cost log is calculated per call from a price table, and hasn't yet been reconciled with those bills.

What I got wrong

• I added complexity without checking it was warranted. Fold began with my own family's needs, so its first features needed no outside evidence. Once I began to see its wider potential, the nice-to-haves went in as well, because at the speed AI builds nothing forced me to descope them; I think that's the real risk of building this fast. Allergies are the clearest example: Fold collected allergy profiles and used them to hide a recipe or show a warning on it, which seemed like a good idea at the time, but I hadn't realised the risks, including legal risks, and I removed allergy profiles on 30 October (decision 10). The Know-how section of technique guides is another; it's a nice-to-have rather than core to the app, but it has been very complex to build well, especially the illustrations. The next point is how I found out.
• The problem statement was never written down whole. I had three partial records of it: decluttering (15 October), cooking mid-flow (16 October) and "cook once, for everyone" (28 October). After about 700 commits in a fortnight, my AI builder team and I wrote a reverse PRD, reconstructing the intent from what had been built. It read those three records as two pivots, which they weren't: it was always one problem, and I've now recorded it whole. More usefully, it found there were no success or stop rules for the beta at all. The response is now a house rule: every new piece of work names the assumption it tests, or it goes to the backlog.
• I broke my own staging rule. The spec said paid photo-to-recipe should wait for evidence from cheaper stages; it was built on day two.
• Free tiers ran out. I won't pay for GitHub or Vercel, so they're hard limits rather than bills. October used 2,163 Actions minutes against 2,000 free, 2,162 of them on CI, and runs stopped on 26 October, and Vercel's 100 deploys a day ran out on 28 October, so nothing could go live for a day. Those limits changed how Fold is built more than the money did: deploys are rationed, docs ride with code, and CI only runs what a change needs.
• A check failed open. For two days the save check crashed and reported "no errors" for every save, and invalid data reached the main branch. Now a crash means "couldn't be checked".
• First use is hard. Three family testers were overwhelmed or missed key features, which is evidence against one of my riskiest assumptions.
• The scheduler toasted risotto rice 49 minutes early, because nothing in the data said it couldn't wait.

The riskiest assumptions, and where they stand

The reverse PRD mapped sixteen assumptions and ranked them by risk and uncertainty. The scores are judgement, not measurement, and they're there to order cheap tests rather than to justify building anything. These matter most:

No claim on this page rests on measured usage yet.

Research before outsiders

Before any household outside my family used Fold, I ran nine Perplexity prompts as desk research, ending with one that tried to prove there was no market. It sharpened the audience (cooks with big cookbook libraries are why people would start; the mixed table is why they'd stay) and the pricing (pay upfront for a first batch, so the expensive part is prepaid). What it couldn't tell me is how many households like mine exist, so a two-part survey is next, with its thresholds agreed before any results arrive.

What's next

An invite-only beta, starting with friends' households and open to anyone who cooks for mixed diets and will tell me how it goes, at fold.emilybuilds.ai, once its safeguards are in place. Three are done: allergy profiles are gone (decision 10), the review panel is merged, with all six reviewers now run (the cost and recipe reviewers for the first time on 1 November), and CI runs on Fold again. The rest aren't yet. The reviews of the beta's accounts work found 13 security issues and 15 privacy issues. The one Critical is moot now that allergy profiles are gone; the rest are waiting for my decisions. The nightly backup isn't set up yet (it failed all three of its recent runs), and I'm setting it up next. The accounts also need trying against a real database rather than stand-ins, and before any household outside my family adds a book, I want a legal view on copyright.

If you're building fast with AI and trying to keep the evidence in step, I'd be glad to compare notes. Get in touch.