First of all, it is absolutely possible to generate 100% of your code if you're working in TypeScript and web apps. I can't speak to other domains, but I know from my own experience that the latest language models can generate beautiful bug-free code end-to-end.

The best models in my experience right now are Claude Sonnet 4.5 and a new undisclosed model called `cheetah` on Cursor, which is next-level quick.

In terms of tools, I prefer using Cursor because I find that I can easily its thoughts and interject when needed. It's also very good at building up context automatically with its semantic code-based search.

It is extremely important to have a good scaffold. You'll notice the scores for SWE Bench are much better when there's a good framework. I've open sourced my monorepo scaffold at the end of this thread.

You basically need to emulate the same process that an expert software engineer would use when coding. 1. Learn (build up context) 2. Make a plan 3. Review the plan 4. Execute on that plan 5. Code review and refactor the results 6. Write tests 7. Run tests, lint, and type check

With custom prompts, and Cursor's features, it's possible to get the AI to closely follow that flow. You can check out the prompt I'm using here, which gets applied to every chat conversation. https://github.com/maccman/ai-...

In general, you want to get out of the way and let the AI do as much as possible. If you are in the loop copying and pasting, then you're just going to slow it down and potentially not give the AI enough context. Cursor has a new "plan" feature that is extremely helpful. I use

I then start building. My favorite LLM is Sonnet 4.5 right now. I will keep an eye on it as it is doing its implementation and read its thoughts to make sure that it's not going off in the wrong direction.

You must understand that the AI works by doing next token prediction, so if you give it a bunch of bad code as context it will generate more bad code. That's why it's very important not to let it go off down a rabbit hole and nip any poor decisions in the bud. Sometimes I will

Once the AI has finished building the feature, I then invoke the `reflect` action, which says: "Is there anything you wish you'd done differently?" This is the most important step. I've found that I get vastly better results the second pass the AI does through the codebase.

In terms of Cursor extensions and MCP tools, I use two: - Context 7, which pulls in documentation for libraries - Perplexity's MCP server (much better than Cursors built-in search. ) Docs here: https://github.com/maccman/ai-...

I use a Mac app called "Spokenly" to dictate everything, so I move at the speed of speech. Right now I'm finding the online GPT-4.0 mini transcribe model the best.

More tips: - Give the AI access to psql for local Postgres querying. Useful for debugging. - Give the AI access to generate and run migrations locally. - Monorepos are back in, they're great for building up AI context. - Make sure your type checking and linting is super fast,

Here is my monorepo scaffold that you can use as a good getting started template. You can check out all the cursor rules I put together as well. https://github.com/maccman/ai-...

Bookmark this thread, I will be adding to it as I learn more.

@maccaw Been using the same workflow to ship all 3 products this year and making money as well. It's a great time for product builders.

🐌 Fibery Q4 2025 Roadmap. Permissions, Files, Mobile... https://community.fibery.io/t/...

UK Has Fallen! An illegal migrant who raped a British child was jailed for 1 year. 3 British men protesting his action were jailed for up to 2 years 4 months. British people must vote the entire political establishment out of every position of power!

As with all the other unskillful relationships I’ve had with various stimuli (alcohol, internet, etc), I was able to skillfully realign my relationship with my phone by dealing with the underlying discomfort / anxiety I was using the phone to cover up. 1

$AAPL - APPLE REPORTEDLY ENTERS ROBOTICS MARKET Apple is expanding smart home production in Vietnam with BYD, making devices like a HomePod-style display and a motorized tabletop robot, expected in 2026–2027. Morgan Stanley says the move brings AI into the physical world and