I went through this exact same panic last month while trying to build a shipping tracker on a shoestring budget. Spent way too much on tokens before I realized the model was just cutting corners to save processing. You might want to consider a very rigid XML structure because if you dont, you will literally throw money away on broken code. Make sure to keep your instructions grounded so it doesnt skip logic to save time.

• wrap your core requirements in tags
• use a 'zero placeholder' policy to stop it from skipping lines
• explicitly force the output to be the full fastapi file I would suggest being very careful with how many tokens you burn. I actually bought a Keychron C3 Pro Mechanical Keyboard specifically because it was cheap and helped me type my own fixes faster when the AI glitched, which saved me from buying more credits. Just watch out for that laziness or it will blow your deadline and your wallet.

TL;DR: Force a scratchpad. In my experience, telling v4 to reason step-by-step before generating full files kills laziness. Tried many variants over the years. ASUS ProArt PA329C 4K Monitor helps.

Just saw this thread and yeah, the scratchpad suggestion is totally the way to go. I would suggest you also watch your token consumption though, especially if you're trying to stay under budget. v4 gets pricey when you demand full files every time. One thing to add is to tell the model to provide complete functional code without any placeholders right at the start. I've noticed it sometimes tries to save space with comments like // your code here, which is super frustrating when you're paying for the output. Be careful about asking for too much at once, as the context window might cut you off mid-file. I usually break the FastAPI routes into smaller chunks to keep things manageable and cheaper. It's a bit tedious but it definitely keeps the logic more consistent for a warehouse app.