How to get useful work out of the new AI Helper — prompting basics for gen builders
Tips & Tricks
Since the model change, I've seen a lot of "the Helper got dumb" posts. Some of that is real — it's a smaller model, and it's worse at guessing what you meant. But here's the part nobody wants to hear: it mostly answers exactly what you ask now, literally. So the fix is on our side. You have to ask better. Here's how, with examples from actual gen building.
- Say what kind of answer you want, not just the topic.
Bad: "What is the data model?" Good: "Explain the data model like I'm five — no code, just the big concept." Also good: "List every field in the data model with one line each saying what it does."
Same topic, totally different answers. The old model guessed which one you wanted. This one won't. Tell it.
- Give it the job, not just a question.
Bad: "Look at my generator." Good: "I'm building a card game gen and the cards overlap on my phone. Here's what I see: [describe]. Fix the layout in index.html and don't touch anything else."
The Helper can edit your code and test it live — but only if you tell it what "done" looks like. "Look at my thing" has no done. "Fix X, don't touch Y" does.
- One job per message.
Bad: "Add a score counter, make it pretty, and also why is the sound not working on iPhones?" Good (message 1): "Add a score counter that goes up by 1 every time the player clicks the target. Show it top-right." Good (message 2, after that works): "Now the sound doesn't play on my iPhone — here's what happens when I tap..."
Three jobs in one message gets you one answer that does none of them well. Stack small wins instead.
- Paste the actual error. All of it.
Bad: "It's broken, it says some error about null." Good: "When I click Generate I get this in the console: [paste the whole thing]. It worked before I added the second list."
The error text is the single most useful thing you can give it. "Something about null" is nothing. Copy-paste is free.
- Say what you already tried and what happened.
Bad: "The image won't load." Good: "The image won't load. It's in src/ as art.png, I'm using , and the preview shows a broken image icon. I re-saved the gen already."
This stops it suggesting the five things you already did, and the details usually contain the actual clue.
- Tell it your constraints up front.
"I'm not a coder, explain every change in plain English." "Most of my visitors are on phones, so test everything small." "Don't change how it looks, only how it works." Constraints you state once save ten messages of corrections. If you build gens for other people, say so — it changes the advice.
- For big builds, make it plan before it touches anything.
"Before you change anything, tell me which files you'll touch and in what order. Wait for me to say go."
This is the highest-leverage trick I know with this model. A wrong plan caught up front costs nothing. A wrong plan caught after twenty edits costs you the afternoon. I make mine do this on anything touching 3+ places.
- When it misses, redirect — don't repeat.
Doesn't work: asking the same question again, but angrier. Works: "That's the parts list, I need the big concept." / "That fix changed the colors too — I only want the layout fixed, put the colors back." / "Too technical, say it again with no code words."
Think of it like directing an intern, not interrogating a database. Point at what's wrong with the answer and say what right looks like.
- Make it verify its own work.
"After you make the change, reload the preview, check the console for errors, and click through the feature yourself before telling me it's done."
The Helper has a live preview with your actual gen in it. A Helper that checks its work catches its own mistakes. One that just types code and says "done" ships bugs. Ask for the check explicitly and you'll feel the difference immediately.
- Keep a standing setup if you build a lot.
If you work on the same gen across many chats, save yourself the preamble: keep a short note (I keep mine in the gen itself) with your rules — how you like answers, your constraints, the stuff you always have to repeat. Paste it at the start of a new chat. Mine tells the Helper to verify everything in the preview, keep answers short, and never touch things I didn't ask about. Thirty seconds of pasting saves the "no, don't do it that way" dance every single session.
The short version: the new Helper is a literalist. Vague in, vague out — except now nothing in the middle rescues you. Specific in, good work out. It's a different skill than before, but it's learnable, and honestly the habits make you better at describing what you want even to humans.