I got 70% of the way with an AI builder and got stuck — help

    Matthew LaCrosse
    2026-07-20
    4 min read
    Getting stuck at 70% is the normal shape of AI-built software, not a sign you did it wrong. The first 70% is screens and happy-path clicks, which AI is great at. The stuck 30% is the hard part it hides: data relationships, auth edge cases, real integrations, and the tool breaking working code each time it regenerates. Diagnose which wall you hit, then bring in a developer for that specific gap — you don't have to start over.

    Almost everyone who builds with an AI tool hits a wall in roughly the same place, and it's weirdly specific: about 70% done. The app looks finished. The screens are there, the buttons click, the demo lands. Then it stops moving — every new thing you ask for breaks two old things, and the AI keeps confidently rewriting code that was already working. You're not bad at this. You've hit the exact seam where AI builders stop feeling like magic.

    Why the wall shows up at 70%

    The first 70% and the last 30% are different kinds of work, and AI is brilliant at one and shaky at the other. The early part is visible and pattern-heavy — layouts, forms, a list that shows some data. There are millions of examples of that, so the model nails it. The last part is invisible and specific to your app: how records actually relate, what happens at the edges, and the rules only your business knows. There are far fewer examples of your exact problem, so the model guesses — and its guesses collide.

    Which wall did you actually hit?

    "Stuck" isn't one problem. It's usually one of a handful, and naming yours tells you what to do next.

    The symptom
    New features keep breaking old ones
    What's really wrong
    The AI regenerates whole files and loses track of what worked
    What it takes to fix
    A developer to freeze the working parts and change code deliberately, not by re-prompting
    The symptom
    Data gets weird once there's real content
    What's really wrong
    The underlying data model is too flat for how your app relates records
    What it takes to fix
    Redesigning the database structure — a foundation fix, not a screen fix
    The symptom
    Login, roles, or permissions half-work
    What's really wrong
    Auth edge cases (resets, sessions, who-can-see-what) the demo never exercised
    What it takes to fix
    Wiring in a real auth provider and testing the messy paths
    The symptom
    The integration won't connect
    What's really wrong
    A real API needs keys, error handling, and server-side code the builder can't fully manage
    What it takes to fix
    Someone writing the glue code and handling the failures properly

    Notice what all four have in common: they're the parts that don't show up in a click-through demo. That's why they surface at 70% — you finally asked the app to do the thing under the hood, and the hood was never really built.

    What to do this week

    1. 1Stop re-prompting the same wall. If three tries haven't fixed it, a fourth won't. Re-generation is what's breaking your working code. Freeze it where it is and change your approach, not your prompt.
    2. 2Write down exactly where it breaks. "When a second user signs up, the first user's data shows up in their account." A precise symptom is worth far more than "it's broken" — it points straight at the real problem.
    3. 3Get a developer to read it for an hour. Not to rebuild — to diagnose. "Is this a foundation problem or a surface one?" That single answer decides whether you patch or rebuild the guts.
    4. 4Fix the foundation before adding anything. If the data model or auth is the wall, no new feature will stick until that's solid. Resist the urge to keep piling on screens over a cracked base.
    5. 5Keep what works as your spec. The 70% you built is a detailed brief for the finished product. Whoever completes it should treat your screens as the answer to "what should this do?" — that's the expensive question, already solved.

    The trap here is sunk cost — "I've come so far, I just need one more prompt." The last 30% is often 70% of the real work, because it's the part that turns the thing into an actual product. Pouring more prompts into a foundation problem is how a weekend project becomes three frustrated months.

    When to bring in a person, and who

    You don't need to throw the app away, and you don't need to hire a whole team. What you need is one competent developer who can look at what the AI built, tell you honestly whether it's a fixable foundation or a rebuild, and then handle the specific 30% you're stuck on. If the guts are sound, that can be a modest, scoped job. If they're not, the rebuild is faster than you fear, because your prototype has already done the hard thinking about what to build.

    If you'd rather hand the whole stuck project to someone who takes it from 70% to launched and stays involved after, that's work we do too. But start with the diagnosis — patch or rebuild — because that call, made honestly, is worth more than any amount of extra prompting.

    Common follow-up questions

    1

    Should I start over from scratch?

    Usually not. Even when the code needs replacing, the prototype you built is a precise specification — it captures the screens, flows, and decisions that are the hard part to figure out. A developer finishing it from that starting point moves faster than the first build did. Start over only if a quick technical review says the foundation is unsalvageable and the design is thin.

    2

    Why does the AI keep breaking things that worked?

    Because many AI builders regenerate whole chunks of code rather than editing surgically, so a request for one change can quietly rewrite parts that were fine. Past a certain complexity, that trade stops paying off. The fix is to have a developer take deliberate control of the code so changes are precise instead of a fresh roll of the dice.

    3

    Is being stuck at 70% a sign my idea is bad?

    No — it's a sign your idea got real enough to need real engineering. The wall shows up precisely when the app has to handle actual data, accounts, and edge cases, which only happens because you built something worth using. The stall is a stage of the work, not a verdict on the idea.

    Want this answered for your exact situation?

    We build and co-found software for people who have the idea and the network but not the technical team. Tell us where you're stuck and we'll give you a straight read — even if the honest answer is "don't build it yet."

    See how we work

    Analytics preferences

    We use Google Analytics to understand which content and services are useful. You can allow or decline optional analytics; essential site functions still work either way.