You can deliver the core of the product - the 80% - in the indicated 20% effort zone, aka a few days.
As for not knowing the code… I actually do read what’s being output. Just because I’m not familiar enough with the language or framework to write it myself with confidence, it doesn’t mean I don’t understand it.
Beyond that, the review loop I described above incorporates experts that are language/framework specific to spot issues. Static analysis can easily spot unused code paths - plus that gets covered by tests too, so any leftover is spotted. And even further, after each MVP cycle (which I use in place of sprints as sprints are for human effort cycles which LLMs don’t really have, so managing actual target goals instead of physical time periods makes more sense), I have a multi-model (as in, different models from different sources, not just e.g. “all Claude but using Sonnet, Opus and Fable”, but more like “Opus 5.5 and GPT 6.1 Sol and [insert currently best model for language]”) quorum of experts review the code, test it end to end, find unit/integration/e2e/UI test gaps, write up everything as big/chore tickets, and have another go of the main loop fix those. Same for documentation, it’s kept up to date in the same loops.
No, it’s not really closer to 60/40.
You can deliver the core of the product - the 80% - in the indicated 20% effort zone, aka a few days.
As for not knowing the code… I actually do read what’s being output. Just because I’m not familiar enough with the language or framework to write it myself with confidence, it doesn’t mean I don’t understand it.
Beyond that, the review loop I described above incorporates experts that are language/framework specific to spot issues. Static analysis can easily spot unused code paths - plus that gets covered by tests too, so any leftover is spotted. And even further, after each MVP cycle (which I use in place of sprints as sprints are for human effort cycles which LLMs don’t really have, so managing actual target goals instead of physical time periods makes more sense), I have a multi-model (as in, different models from different sources, not just e.g. “all Claude but using Sonnet, Opus and Fable”, but more like “Opus 5.5 and GPT 6.1 Sol and [insert currently best model for language]”) quorum of experts review the code, test it end to end, find unit/integration/e2e/UI test gaps, write up everything as big/chore tickets, and have another go of the main loop fix those. Same for documentation, it’s kept up to date in the same loops.
You get 80% there by being a software engineer to start with, without proper discipline and experience in the field it won’t be 80/20.