I’ll write a bit about my experience using LLMs after two years, so to speak.

I’ll write a bit about my experience using LLMs after two years, so to speak.
First off, what I usually do:
- throw together UX prototypes as a front/flows and all that — designing features like this conveys the essence of what you want to build better
- make pitch decks, business models, and all that kind of crap
- write lots of meaningful emails like: so why the hell do you need this feature and where will it lead, or essays on the theme of “strategy is everything, and I don’t give a shit what we ship tomorrow”
- think a lot, meet with people, and write follow-ups

Tools:
- ChatGPT
- Claude code/cursor
- advanced voice mode ChatGPT

For myself I figured out a few commandments / rules for how to succeed working with LLMs:
- well if you started doing something and you see it went somewhere wrong or forgot some important details — I start over. Nothing will help fix or add something into the already existing “stream of consciousness” of the model
- it’s extremely important to formulate for yourself what result you want (a site similar to X, an Excel with these fields and these dependencies) ahead of time, and even better to understand how you yourself would do it by hand. Then everything goes way nicer. Like, I ask it to generate Excels while specifying it needs to write formulas, not just numbers in a table. Then it comes out better.
- vibe coding is a pretty tool. But if I’ve never written code for “something” before, I won’t assemble anything besides a website. Because in the process you need to understand the details and sometimes the essence of the errors, and they will happen. If you don’t understand the essence of the errors — the sheet of changes will be out of control. Then see p1 (and you don’t even understand that something went wrong)
- but the real top is to dictate thoughts right after a meeting in voice mode and iterate by voice too — it lets you very quickly formulate the main outcomes and write a balanced analysis for the team or as a reply

Personally my efficiency has increased by multiples. I used to be really lazy about building models and calculations by hand, and now I do basic code for something or an Excel super fast. 50x speedup. Since my main difficulty was that without constant use I quickly forget the tools (Excel, front-end, or complex calculations), I just take forever typing in texts and formulas and constantly screw up syntax. Obviously I edit texts with the history in mind and often dump all versions of decks or emails into a folder/project in ChatGPT.

TLDR: for me this is still far from automation. And honestly I don’t believe this threshold can be stepped over with the current generation of the tech. It’s super cool. But it feels like an electric car from
Detroit Electric. (https://ru.wikipedia.org/wiki/Detroit_Electric)

P.S. haven’t figured out why the hell to generate video and pictures yet