they are saying the prompt is the source, the code is the artifact. I think it is a valid comparison. when reviewing a PR, I do want to see the prompt(s) that generated the change.
My latest feature was an overarching change across 4 different stacks, about a full work day of going back and forth iterating on different designs, generating workarounds for stack/language specific quirks and the business' usage patterns. Multiple fresh agent sessions, review passes, exploratory work with subagents, etc. "The prompt(s) that generated the change" in that context is multiple hours of discussion, Q&A, refactoring and feedback. Is that really the value, pages and pages of reactive prompting?
Yes. Because that is where the intent is/was. In the future, as more changes are made, having the history of the intent is helpful context. And in that back and forth there are also details of what you didn't want, and why. Some of that can be captured in decision docs, etc.
when devs work in isolation, that context/memory is siloed. Same is true when working across platforms (codex, claude, etc).
What puzzles me is, what matters is context. Most of the context would be the specific model I used and its training data, the result of tool calls, and the model's inner monologue/thinking (which is increasingly getting encrypted by frontier models). My prompts are a tiny sliver of that, and most of them dependended on the actual answers I was getting from the model in the middle of the whole process. Or the result of handoffs, or internal compactions.
What's the value of getting all my prompts if the rest of the equation is unknowable? Unless you mean the value is in the full agent log, but even then, frontier models and harnesses would still be hiding a major part of the equation from us.
I'm not sure what level of detail those prompts you write about have so I might misunderstand you.
Before AI we had prompts called "requirements" and of course they were the starting point for reviewing code: the customer needs X, the code doing that is Y.
With agents that prompt became much longer, with the original requirement a short note at the beginning of the conversation with the AI, but the resulting code is about the same size. Maybe in YOLO development the original requirement is all the AI needs but I'll be surprised because requirements from customers are too fuzzy and can be turned into very different features, often too different from what the customers had in mind.
Actually I never looked at the detailed prompts of coworkers, only at the original requirements from customers. How they instructed their agents is not particularly interesting except maybe to get a gist of the different approaches.
agents are pretty good at cleaning up a codebase, finding dead code, fixing bad abstractions, etc. You just have to spend some focusing the agents on that goal.
my cellphone has been named “sneakernet” for years. it’s a throwback to a time when it was faster to walk a zip disk across campus than it was to send it.
I'd walk with my CRT monitor and tower (stuffed all cables, mouse, keyboard inside) over to my friend's to play games/mess with computers a couple decades ago. We called it the "sneakernet", too.
Fun fact. My wife and I met on something similar in 2006 or so. "My Blog Log" was a sidebar widget that showed other people who we reading that blog. She was in marketing, I was in tech, and it was a blog at that intersection written by Rex Hammock (RIP).
i have been thinking about this from a different direction: how do we make these shared within a company in a way that increases the productivity floor of the team/department/company. Sure, they can still be extended/enhanced by individuals, but we don’t need everyone configuring mcps, building institutional memory, etc.
for me, it’s not about the cost to leave, it’s about lowering the cost of onboarding and change.
