THIS WEEK’S STORY
What All Fast Builders Have in Common (spoiler: it’s not talent)
I keep noticing the same pattern with AI builders. Some people progress insanely fast with zero coding experience, while others stay stuck for months.
The difference isn’t intelligence. It’s how they talk to AI.
Most non devs think they can just open Lovable, Cursor, or v0, type a few prompts like “build me a full app,” and expect everything to magically work with no errors. Which is kind of ironic when you think about it.
You might actually get a beautiful frontend with the features you asked for, but then reality hits. The database is not connected. API keys are exposed on the client side so anyone can steal them. Subscriptions are not tied to users. Auth is broken. “Simple features” don’t behave the way you assumed.
This doesn’t happen because the AI did a bad job at writing code. In most cases it did exactly what you asked, and did its best to fill the gaps. The problem is that you didn’t direct it properly.
Here’s the real kicker: architecture thinking is the new coding.
To get good at it, you need to start seeing your project like a LEGO build. Your job is to write the instructions manual (your PRD doc) and then assemble the project step by step. That means using different LLMs like ChatGPT, Grok, Perplexity, Claude to research tools, frameworks, APIs, tech stacks, and understand how all the pieces should work together.
AI is only as good as the architecture you describe.
Once you’re confident in your instructions manual, then you can start “building the LEGO.” An API brick here for feature X. A UI component there. Stripe plugged in for payments. Auth wired up properly. One piece at a time.
Fast builders do exactly this. They write PRDs, break tasks cleanly, ask AI to explain what it just did, commit code often, and treat AI like a collaborator instead of a genie.
Start thinking about AI coding this way and the ceiling disappears.
HEADLINES
Anthorpic Interviewer release
Interviewer is a new system that uses Claude to run live, adaptive interviews with real workers so researchers can understand how AI is actually used in day-to-day workflows. Instead of relying on chat logs, Interviewer plans questions, adjusts in real time, and analyzes patterns across professions.
Their first batch of 1,250 interviews revealed a few big insights:
Workers automate routine tasks but keep the responsibilities tied to their identity
Creatives report strong productivity gains but still manage stigma around using AI
Scientists use AI heavily for coding and writing while avoiding hypothesis-level decisions
91 percent of scientists want deeper AI research assistance
You can join future interviews directly on Claude.
THIS WEEK’S TIP
My Cursor workflow for building apps
I’ve been developing with Cursor for about 1.5 years now and saw every single update and how it improved my workflow. Today, I think I reached a pretty stable process.
Here are the rules I live by:
You must have a PRD doc – the first step when starting a new project or taking over someone’s else is to create a prd.md doc documenting all aspects of the project: architecture, tech stack, user flow, UI/UX, backend logic etc. This is extremely important because it gives general context to your ai agents when building new features. This prevents the LLMs from hallucinating a new tech stack elements (for example Express vs Nextjs api routes) or duplicating files.
Isolate agent requests – as soon as you’re done with an implementation, “fire” that agent right away and start a new chat. This avoids context overload and ultimately your agent going off the rails. For instance, spin up a AI chat to build a pricing page UI wireframe, another chat to perfect the design, and another chat for each individual backend integration (with Stripe for instance).
Use “plan mode” for complex requests – when working on a big feature, select the “Plan” mode and select a thinking model like Opus 4.5 or GPT 5.1 Codex. Start by describing in detail what you want to achieve, highlighting all the relevant files to read as well as external documentation URLs. Then finish the prompt with “ask me any question, doubt or clarification you need before writing your plan”. The plan mode will give you multiple choice questions for you to answer to make sure it really understands what you wanna build. When you are ready to build, approve the plan by clicking “build” and select either “auto” or “composer 1” to write the code (cheaper and faster once the plan is detailed enough)
Ask questions about your code – not sure about what the agent implemented? Toggle the “ask” mode and tell it to explain the recent implementation in plain english. It will make your life s much easier when debugging or implementing new feature later.
Use the built-in browser – Cursor now has a built in browser where you can visualize anything from a website to the localhost. But the coolest feature is that you are able to select elements from the browser and reference them directly in the AI chat. This streamlines massively iterations and allows you to be more precise when giving context to the agent
Always commit your code (GitHub) – I can’t stress how important this is, especially for less technical people who have all the code generated by AI. Whenever you have a “stable” version of your code after testing it, you should commit and push the changes to github. This means that your “stable” code lives in a cloud repository where you can always revert back to in case the agent start hallucinating and breaking features. Also, you should use “branches” when working on new implementations so you don’t touch the main version of your code. I will do a more in depth tutorial for non devs in the future about this.
There are many more but these are some important ones. If you need help with any, drop me a line. And make sure to read next weeks’ posts too for more tips 😉
QUICK HITS
AWS launches Trainium3 chip: 4x faster, 40% more efficient than Trainium2, targeting Nvidia's dominance (AWS re:Invent Keynote)
Gemini 3 "Deep Think" mode rolls out with advanced reasoning; outperforms most models on complex coding benchmarks (Google Blog)
Cursor overtakes GitHub Copilot as fastest-growing AI code editor, now used by 2M+ devs monthly (Cursor State of AI Coding Report)
Anthropic acquires Bun to supercharge Claude's code generation speed and runtime performance (Anthropic Announcement)
Gen AI platforms hit 7 billion monthly visits, now rivaling TikTok and Instagram in web traffic (Similarweb Dec 2025 Report)
And That is all for this week!
See you next Saturday!
Filippo
Filippo Pietrantonio
Whenever you are ready, here are 3 ways I can help you:
Become a member of the aiOS App builders community: In this group, I share content that you won’t find anywhere else + my bullet-proof workflows to build mobile apps with AI to escape your 9 to 5 and launch your dream app.
Get personal help and guidance from me: If you have a great app idea, you are eager to launch (timing is everything!), and you’ve tried different methods but still find yourself stuck with budget constraints… I can help you se yourself free and create what you truly want. Click here and let’s see how we can work together.
