My LLM coding workflow going into 2026

Useful pointers, which match my experience well when trying this out. The one big question that springs to my mind is: when you've done all these steps, how much time did you really save?

You've chosen a strategy, broke down the solution into small easy to code parts, validated business logic, noted traps to avoid, searched for all relevant code, set up a context packing document particular to this section of the code base, experimented with multiple agents, reviewed each version of the code to see if you understand it-

That sounds like a lot of work!

And why, so the AI can do the last 10% of actually coding it up? Is there even a speedup here over doing it yourself? There's some evidence against AI speedups [1]. Of course some of these steps are themselves AI enhanced, and some of them are part of the work regardless of whether you use AI.

I still feel skeptical on this workflow and how big the gains could possibly be. I feel more for the alternate approach of writing it yourself, but using the AI for the boring parts (e.g. copy this section, but use those functions), or for sparring / research. I have however no data to show which approach takes less time.

I can however tell you from experimenting with full on vibe coding- you can do it with half the attention you'd give the task yourself. Is that what I'm gaining, that I can read a book during code generation? (https://xkcd.com/303/)

[1] https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...