We gotta move fast and we are seeing the generated slop day by day. If you are lucky and don't see it, then please don't participate in this conversation. We don't need to be belittled down. We know where that path leads too.
However, for those who have to read 2000 lines of code every single hour, (2000 is exaggeration. But at the end of the PR, it adds up to be 2000.)
1. how are you managing the mental bandwidth?
2. How are you verifying these code and signing them off?
3. Any tips, tricks, ideas that will keep us relevant here?
The biggest concern is understanding the maths and research code behind it. The shitty abstraction and duplication just puts off reading the code.
Regarding context collapse, the model doesn't have any idea of across different sessions. The token inefficiency is off the charts.
4. The problem is that what are we going to do about this token inefficiency? I find myself rewriting same instruction because the fix from last patch broke the system so this patch it needs to write a new fix, then proceed with feature implementation. The same command, function, calls, doesn't work across two session.
5. TDD is not going to help when half of the tests are meaningless?
tldr; Problem is not that we can code by hand, problem is there is no way to code this sheer volume (feature or design, sometimes entire product) by hand in the amount of time that we have. That is the problem.
However, for those who have to read 2000 lines of code every single hour, (2000 is exaggeration. But at the end of the PR, it adds up to be 2000.)
1. how are you managing the mental bandwidth?
2. How are you verifying these code and signing them off?
3. Any tips, tricks, ideas that will keep us relevant here?
The biggest concern is understanding the maths and research code behind it. The shitty abstraction and duplication just puts off reading the code.
Regarding context collapse, the model doesn't have any idea of across different sessions. The token inefficiency is off the charts.
4. The problem is that what are we going to do about this token inefficiency? I find myself rewriting same instruction because the fix from last patch broke the system so this patch it needs to write a new fix, then proceed with feature implementation. The same command, function, calls, doesn't work across two session.
5. TDD is not going to help when half of the tests are meaningless?
tldr; Problem is not that we can code by hand, problem is there is no way to code this sheer volume (feature or design, sometimes entire product) by hand in the amount of time that we have. That is the problem.
why are you even attempting to read 2000 loc per hour? that won't work.
if ai generated the code, ai can review it. I can review perhaps pieces that are deemed priority. manual and automated testing are both required.
for the rest, please number your Qs so they're easy to address.
But how are you scoping the features these days? Unlike back in the day, these days a feature is equivalent of research paper
if you prompt differently, you can instruct them to work on small things
they have a strong tendency to over engineer (volume), prompt against this - in permanent files like AGENTS.md or per-session in follow on messages
Yeah this over engineering is really annoying bit
Do you really think that's what's happening?
staying in the loop and being methodical about it (at least trying)
I more often find the walls of markdown to be more onerous than the code changes
would you mind sharing how you reduce it. I am testing hacks like asking to draw in block diagrams, write like RFC, write like algo and pseudocode
context engineering, with guidelines drawn from examples in my own projects
more up front planning, in a markdown file, multiple sessions
review processes, for the plan and then the code
the "recursive self improvement" (at the harness/context engineering level) feels like it is starting to kick in