Post #1715486

2026-04-27 01:07 UTC

@brandonscript@appdot.net @timbray@cosocial.ca I'd like to see a setup where I still do the work of coding and the AI is solely a pairing partner. Delegating the entire process to an LLM including the actual construction removes all of my agency from the process, and delegating small decisions to LLM eventually leads to delegating all of them. If you don't exercise your critical thinking abilities you will lose them. Too many developers are choosing that path.

Replies (4)

@francois@ruby.social 2026-04-27 01:27

@headius@mastodon.social I did that by doing ping- pong with the LLM. The LLM writes the failing test case. I make it pass, and make many decisions along the way, some invalidating the test. Then we switch roles: I write the next failing test case, and it brings us back to green. This works because before, I’ve planned things out in great details, with the LLM’s help. I’ve taken many big decisions up-front. It works, but the LLM must populate its context, so the first few turns are slow. Then, it gets better.

Open ##1715485
@pointlessone@status.pointless.one 2026-04-28 23:08

@headius@mastodon.social I had some limited success using an agent in Ask mode with a new to me codebase. It can go read a bunch of files faster than me and can give relatively OK answers to questions like where can I find auth code or what does this big undocumented class do. @brandonscript@appdot.net @timbray@cosocial.ca @adarsh@ruby.social

Open ##1758661
@brandonscript@appdot.net 2026-04-27 03:08

@headius@mastodon.social yep. That's how I do it. It's my rubber duck, and it does all the tedium so I don't have to @timbray@cosocial.ca

Open ##2189721
@mnl@hachyderm.io 2026-05-04 12:57

@headius@mastodon.social @brandonscript@appdot.net @timbray@cosocial.ca I literally split the work physically. I do most of my work in my sketchbook, and have the agents code it up and write a “diary” (prompting trick to get more thoughtful and reviewable documents to review than a dump of diffs and slop language) / write documents about the codebase and potential implementation steps (they are explicitly not “plans” but “analysis and design guides”), then push those back to the tablet and work on them away from the computer. The UX of llm-assisted tools as “chat with a bot in a terminal that scrolls by too fast to read” is pretty antithetical to deep thinking, imo. https://the.scapegoat.dev/slowing-down-in-the-age-of-coding-agents/ It’s been pretty great so far to do 90% of my work away from the computer.

Open ##2189723