IMHO neither deep codebase familiarity nor communication skills are durable advantages. Arguably they already aren't. Claudish is a problem for Claude but I've seen no signs of equivalent problems with GPT 5.6. It communicates clearly. The ability to find things and use the existing codebase well is largely dependent on how well structured it already is. I've mostly used AIs with codebases that I wrote myself and they're well structured. Some have plenty of internal dev-facing documentation too. I'm often surprised at how well the models use the internal abstractions - easily as well as I would have. Except over time I'll forget the details of codebases I don't work deeply on, whereas the model is rediscovering each time by using its far superior reading speed, so its ability to use the details won't degrade and mine will.

I currently do add value by guiding the models when they overlook a better way to do things. But it feels like writing code by hand is going to go the same way as writing assembly language by hand already did. There will be rare cases where it's necessary for some reason, but coding will steadily become seen as some sort of dark wizardry only a handful of old codgers know. Eventually the art will be lost entirely, a bit like how military archery and spoken Latin were skills Roman warriors took for granted but today ~nobody can do them.

So where's the durable advantage? I see a few candidates:

1. AI bubble pops and the rate at which the big model firms ship features takes a dive. Working around model/tooling limitations and optimizing costs ends up becoming a source of sustained value.

2. Business change consulting. Models are passive, so you need to think of a question to ask. Someone actively using their technical knowledge to scout out opportunities and come up with creative solutions might have value for a while, not because AI couldn't come up with these ideas given the right prompt but because it just won't be given the right prompt without you.

3. Building "AI native" companies. Big technology transitions often leave existing institutions behind because they can't adapt, either culturally or in terms of skills. The internet was one example of that - companies like Amazon should theoretically have had no chance given the existence of many highly competitive retail firms - and the inability of Swiss watch companies to deliver smartwatches is another specific example. So it's possible that AI will be like that and fully absorbing the consequences won't be possible for existing firms with large labour forces, opening up an opportunity for a wide range of new startups to come in and take shares of mature markets.

None of these are coding.