20% knowledge could be enough to fix errors yourself. E.g. this function name has a typo, or this calculation looks wrong, or this code has been repeated verbatim 40 times, maybe it should be a function, or why did the LLM change this totally unrelated area of the code?

If you only take exactly what the LLM spits out you can only verify by running. If it is broken in edge cases you can only prompt the LLM to fix. You then have to run the code again to verify the fix, and with zero understanding it may have broken some other edge case (LLMs do this all the time especially overly eager reasoning models).