News for hackers. Stuff that matters.
Re:"I can tell you wrote this yourself" by mrwh on Friday November 20, 2026 @09:09AM (#50574886)
An AI design doc has all the outward appearances of a -- good! excellent sometimes! -- preAI design doc, in my experience. But it also has no reason to exist. There's no social aspect, since no one has gone through the process of actually understanding the systems involved, there's no meaningful way to review it ("why do things this way?" "Because it said to do things this way"). And it's not needed for the implementation either, for which the original prompt, whatever it was, suffices. By industrialising the output of a process, we forget why it was necessary to go to the trouble in creating it by hand in the first place. Perhaps that's ultimately how things advance, and right now we're in the absurd middle, treating proofs of work like the work itself.
Re:"I can tell you wrote this yourself" by mceachen on Friday November 20, 2026 @10:38AM (#50576316)
> it also has no reason to exist.

I despised having to produce design documents for software projects 20 years ago (the last time I had to do it as a software engineer). I wanted to jump straight to coding. I begrudgingly found that, much like research papers, it's a forcing function in really understanding the problem space.

Over the last year, I've found that, at least for nontrivial things, if my agent makes a terse (under 300 line) technical design doc that describes their research and their tasks, it can highlight any disconnects between what I asked and the agent's interpretation. It also serves as context to ramp up cross-model reviews, and as historical reference material when I have to Chesterson's fence something later. I wrote up more about this here: https://photostructure.com/coding/claude-code-tpp/

Re:"I can tell you wrote this yourself" by shermantanktop on Friday November 20, 2026 @10:53AM (#50576587)
I know a lot of engineers who hate writing design docs. I think of them as being more comfortable with ideas in code than with English. And there are good reasons for that. But ideas in code are trapped, whereas ideas in English are portable and potentially viral in how they move between people.

But all those English-hating engineers now produce shitty AI design docs, which tick a box but have less than zero influence on anyone else’s thinking.

Re:"I can tell you wrote this yourself" by techmunky on Friday November 20, 2026 @12:05PM (#50578035)
If you're using your LLMs to create the product A - Z then sure these docs don't matter. But assuming you or a team is driving the product and engineering architecture then these docs do matter. I still build out docs because they are a reviewable source of truth, a reliable historical knowledge sink, and because I still know how my product works.