As I work more with AI, I’ve came to the conclusion that I have no patience to read AI-generated content, whether the content is right or wrong. I just feel like it’s time wasted. Countless of examples: meeting summaries (nobody reads them), auto generated code (we usually do it for prototypes and pocs, if it works, we ship it, no reviews. For serious stuff we take care of the code carefully), and a large etc.
I like AI on the producing side. Not so much on the consuming side.
For me, AI meeting summaries are pretty useful. The only way I see they're not useful for you is that you're disciplined enough to write down a plan based on the meeting subject.
- I'm reviewing the last meeting of a regular meeting cadence to see what we need to discuss.
- I put it in a lookup (vector store, whatever) so I can do things like "what was the thing customer xyz said they needed to integrate against".
Those are pretty useful. But I don't usually read the whole meeting notes.
I think this is probably more broadly true too. AI can generate far more text than we can process, and text treatises on what an AI was prompted to say is pretty useless. But generating text not with the purpose of presenting it to the user but as a cold store of information that can be paired with good retrieval can be pretty useful.
Im in a large company, sometimes we have long incident meetings running for hours and new idiots join in the middle "what happened?". Now at least we can get summaries of the past hours during the meeting to catch up without bothering everyone !
While in principle that should be great I don't even slightly trust it as a technique because you're compounding points at which the LLM can get things wrong. First you've got the speech to text engine, which will introduce errors based on things like people mumbling, or a bird shouting outside the window. That's then fed into a summarising LLM to make the meeting notes, which may latch onto the errors in the speech to text engine, or just make up its own new and exciting misinterpretations. Finally you're feeding those into some sort of document store to ask another LLM questions about them, and that LLM too can misinterpret things in interesting ways. Its like playing a game of chinese whispers with yourself.
Understand your concerns; I'm interested in trying it to see if I get value from it (things don't have to be perfect to be useful). For the last part, you can ask it to include a source for how it answers questions and check that.
I don't read most of my (work) emails either, but I think the most important part of AI generated meeting notes is that they're searchable / indexable, in the off chance that you do need to find or refer to something mentioned in an article.
But to be blunt / irreverent, it's the same with Git commit messages or technical documentation; nobody reads them unless they need them, and only the bits that are important to them at that point in time.
I can't help but see the irony of complaining that people don't read git commit messages or technical documentation in the comments for a product to assist in code review.
You know what really, really, helps while doing code review? Good commit messages, and more generally, good commit practices so that each commit is describing a set of changes which make sense together. If you have that then code review becomes much easier, you just step through each commit in turn and you can see how the code got to be where it is now, rather than Github's default "here's everything, good luck" view.
The other thing that helps? Technical documentation that describes why things are as they are, and what we're trying to achieve with a piece of work.
That's fair! If there were a "minimal" mode where you could still access callers, data flows, and dependencies with no AI text, would it be helpful for your reviews?
Not parent, but in my opinion the answer here is yes. I agree that there is a real need here and a potentially solid value proposition (which is not the case with a lot of vscode-fork+LLM-based starups) but the whole point should be to combat the verbosity and featurelessness of LLM-generated code and text. Using an LLM on the backend to discover meaningful connections in the codebase may sometimes be the right call but the output of that analysis should be some simple visual indication of control flow or dependency like you mention. At a first look the output in the editor looks more like an expansion rather than a distillation.
Unrelated, but I don't know why I expected the website and editor theme to be hay-yellow and or hay-yellow and black instead of the classic purple on black :)
Thanks for the opinion! That makes a lot of sense and I like the concept of being an extension of a user's own analysis vs hosing them with information.
Yeah originally I thought of using yellow/brown or yellow/black but for some reason I didn't like the color. Plenty of time to go back though!
honestly i feel the same way and i can't quite put into words why. I guess if I had to -- I think it's because I know not all AI generated stuff is equally created and that some people are terrible at prompting/or don't even proofread the stuff that's outputted, so I have this internal barometer that screams "you're likely wasting your time reading this" and so I just learned to avoid it entirely. Which is sad, because clearly now a ton of stuff is AI generated, so I barely read anything, _especially_ if I see any signals like "it's not just this, it's that"
Products like these make me realize we're solving for the wrong problems with a lot of these AI solutions. I don't want you to take this as a hit to you or your product, I actually think it's extremely cool and will likely find a use. But from my perspective if this is a product you think you need, then you likely have a bigger organizational issue, as PRs are probably the last thing that I would want an AI 'intern' to organize for me.
Could you expound on this? In my experience as a software engineer, a pull request could fall into one of two buckets (assuming it's not trivial):
1. The PR is not organized by the author so it's skimmed and not fully understood because it's so hard to follow along
2. The PR author puts a lot of time into organizing the pull request (crafting each commit, trying to build a narrative, etc.) and the review is thorough, but still not easy
I think organization helps the 1st case and obviates the need for the author to spend so much time crafting the PR in the 2nd case (and eliminates messy updates that need to be carefully slotted in).
This is where I feel like we've solved a third-order problem. If you're sorting all PRs into those two buckets then you should probably take a step back and redefine what a PR is for your organization, as both 1 and 2 make the assumption that the PR is too big to review in a single sit down or that the author didn't put in enough effort to craft their PR. Both of these should just be rejected outright in favor of doing things in a smaller more manageable way, instead of having an AI sort through something that a human should have started with. Obviously this is more of an ideal situation and a lot of companies don't work on the ideal which is why I think your product will find good use because companies don't want to invest in slowing down, only going faster.
> you should probably take a step back and redefine what a PR is for your organization
I agree with this wholeheartedly if you are in a role that allows you to redefine what a PR is. In almost every organization that I've worked for, the PR is defined several levels above my pay grade and suggesting changes/updates/etc is usually seen as complaining.
Interesting. At my previous company there was a debate about smaller PRs vs bigger PRs and the end conclusion was that there are tradeoffs in being able to deal with 2-5 bite-sized PRs vs one large PR. The largest one being that it's hard to grasp the totality of the pull request and how the different PRs work together.
> companies don't want to invest in slowing down, only going faster.
I do think this is the way things are going to go moving forward, for better or for worse!
My solution is to organize my PRs as a sequence of commits that explain what they do, and then a PR description that gives an overview and motivates the changes. I've gotten really positive feedback on this, and it dramatically speeds up reviews of my code. Overall less work for the team. (And it often helps me find problems before I even submit the PR.)
As for other people's PRs? If they don't give a good summary, I ask them to write one.
> If they don't give a good summary, I ask them to write one.
Exactly; if people can't be bothered to describe (and justify) their work, or if they outsource it to AI that creates something overly wordy and possibly wrong, why should I be bothered to review it?
How much time does it take you to craft the PR compared to writing the code? Just curious if crafting the PR becomes significantly easier with practice.
I think this is a valid part of the "crafting PR" skill that's under appreciated, and part of the goal of Haystack here is to make that part of PR craft effortless.
Very glad you're working on this! Here is my wish list as a code reviewer:
1. Allow me to step through the code execution paths that have been modified in the pull request, based on the tests that have been modified.
2. Allow me to see the data being handled in variables as I look through the code.
3. Allow me to see code coverage of each part of the code.
4. Show me the full file as I am navigating through the program execution so that I can feel the level of abstraction and notice nearby repetition or code that would benefit from being cleaned up.
> 1. Allow me to step through the code execution paths that have been modified in the pull request, based on the tests that have been modified.
Not sure if I fully grasp this! We tried to kind of do this in previous iterations (show call graphs all at once) and it gets messy very fast. Could you elaborate on this point in particular?
Sure - imagine my PR adds one new test which test one new function.
Starting from the test, allow me to step through the program execution, just like a debugger, to observe variables, surrounding code, and the complete file.
If you read only the covered lines of code in a linear way, you'd miss the refactoring opportunities because you aren't looking at the rest of the file.
This nails a real problem. Non-trivial PRs need two passes: first grok the entrypoints and touched files to grasp the conceptual change and review order, then dive into each block of changes with context.
It could be just fatigue over the technology itself, but “like you wrote them yourself” sounds too much like a dog whistle for a user base of programmers working with AI generated PRs. Which means that this would be AI iterating over AI.
Feedback: Try speeding up your demo animations and resize the mouse to its regular size. My estimate is that if the marketing copy explains what a thing is, what it does and why it’s useful then all a visitor wants to see in an image is things go pop, boom and whoosh.
Not sure if I fully grasp what you mean by dog whistling, but at the end of the day, like another commenter said, Haystack is also pretty helpful for when you're done experimenting with a piece of work and need to see what an AI has generated.
I love this idea, trying it out now. There is QUITE a delay doing the analysis, which is reasonable, so I assume as a productionized (non-demo) release this will be async?
There's a GitHub app that you can install on your repo.
If you install and subscribe to the product, we create a link for you every time you make a pull request. We're working (literally right now!) on making it create a link every time you're assigned a review as well.
We'll also speed up the time in the future (it's pretty slow)!
I would really want to use this, maybe about once a week, for major PRs. I find it absurd that we all get AI help writing large features but very little help when doing the approx same job in reviewing that code. I actually would even read my own PRs with it, as my workflow with AI is to prompt it to acheive building some feature/goal, then only review the code once things work (this is an oversimplification).
Yes! I personally iterate without understanding the code too much (since it'll change), and fully assess what Claude Code (in my case) has done after I finished a piece of work.
I think tools like this are useful, but they can never replace the quality of the narrative that someone who actually wrote the code can come up with.
There's just so much contextual data outside of the code itself that you miss out on. This looks like an improvement over Github Co-Pilot generated summaries, but that's not hard.
This is a cool problem to solve! especially as more and more code is being generated with AI. But are you pivoting away entirely from code visualization? Curious to hear why it didn't work out.
We may go back to code visualization for editors, but after testing it with a lot of engineers, we found that it was too confusing for code reviews. The main issue was just how alien the interface is!
1. Pricing would be $20 per person and we'd spin up an analysis for every PR you create/are assigned to review
2. We don't train or retain anything related to your codebase. We do send all the diffs to Open AI and/or Anthropic (and we have agentic grepping so it can see other parts of the codebase as well)
3. Do you mean the ability to run this on your own servers (or even on your own computer with local LLMs)? We do have plans for the former, but I don't know how big the lift on the latter will be so I'm not sure!
> We don't train or retain anything related to your codebase.
It would be good to have this mentioned on the website somewhere, as part of a privacy policy. Right now I can't find any details on the site, which is preventing me from trying this out with a production repository.
We use a mixture of LLMs (e.g. to identify the important variables to trace), language servers (to understand where a variable is originally defined and where it's used).
Pretty cool. Not from a big firm, but I really like the idea of adding metadata to PRs!
One counterintuitive thing about navigation is that I keep hitting the Back button, hoping to go back to `Pull request summary`, which feels like the main navigation page, but nothing happens. You already implement navigation back and forward in history, so why not do it in the browser router too?
Sorry for the confusion! Missing browser navigation is the problem. The virtual back buttons you put in the top left are working as I'd expect browser nav to do. I keep trying to go back, it would feel so natural.
As I work more with AI, I’ve came to the conclusion that I have no patience to read AI-generated content, whether the content is right or wrong. I just feel like it’s time wasted. Countless of examples: meeting summaries (nobody reads them), auto generated code (we usually do it for prototypes and pocs, if it works, we ship it, no reviews. For serious stuff we take care of the code carefully), and a large etc.
I like AI on the producing side. Not so much on the consuming side.
For me, AI meeting summaries are pretty useful. The only way I see they're not useful for you is that you're disciplined enough to write down a plan based on the meeting subject.
I tend to agree. Except if it's text generated by me for me.
I don't want you to send me a AI-generated summary of anything, but if I initiated it looking for answers, then it's much more helpful.
I'm not doing this much now, but this AI-generated text might be more useful if you use AI to ask questions using it as a source.
Meeting notes are useful in two ways, for me:
- I'm reviewing the last meeting of a regular meeting cadence to see what we need to discuss.
- I put it in a lookup (vector store, whatever) so I can do things like "what was the thing customer xyz said they needed to integrate against".
Those are pretty useful. But I don't usually read the whole meeting notes.
I think this is probably more broadly true too. AI can generate far more text than we can process, and text treatises on what an AI was prompted to say is pretty useless. But generating text not with the purpose of presenting it to the user but as a cold store of information that can be paired with good retrieval can be pretty useful.
Im in a large company, sometimes we have long incident meetings running for hours and new idiots join in the middle "what happened?". Now at least we can get summaries of the past hours during the meeting to catch up without bothering everyone !
I'm aware this idealistic, but I would suggest simply not doing that. Don't have hours long meetings with people joining midway through.
Seems very idealistic :)
I think this is about when the app is broken and people are keeping a meeting app open to communicate with each other as they scramble to fix things.
So the limitation here is more about problems not being solved yet rather than how a 'meeting' is organized.
While in principle that should be great I don't even slightly trust it as a technique because you're compounding points at which the LLM can get things wrong. First you've got the speech to text engine, which will introduce errors based on things like people mumbling, or a bird shouting outside the window. That's then fed into a summarising LLM to make the meeting notes, which may latch onto the errors in the speech to text engine, or just make up its own new and exciting misinterpretations. Finally you're feeding those into some sort of document store to ask another LLM questions about them, and that LLM too can misinterpret things in interesting ways. Its like playing a game of chinese whispers with yourself.
Understand your concerns; I'm interested in trying it to see if I get value from it (things don't have to be perfect to be useful). For the last part, you can ask it to include a source for how it answers questions and check that.
I don't read most of my (work) emails either, but I think the most important part of AI generated meeting notes is that they're searchable / indexable, in the off chance that you do need to find or refer to something mentioned in an article.
But to be blunt / irreverent, it's the same with Git commit messages or technical documentation; nobody reads them unless they need them, and only the bits that are important to them at that point in time.
I can't help but see the irony of complaining that people don't read git commit messages or technical documentation in the comments for a product to assist in code review.
You know what really, really, helps while doing code review? Good commit messages, and more generally, good commit practices so that each commit is describing a set of changes which make sense together. If you have that then code review becomes much easier, you just step through each commit in turn and you can see how the code got to be where it is now, rather than Github's default "here's everything, good luck" view.
The other thing that helps? Technical documentation that describes why things are as they are, and what we're trying to achieve with a piece of work.
Maybe AI doing rebase, code chunking and commits will help? This kinda makes sense. Reviewable+mergables chunks do make the code way faster to merge.
That's fair! If there were a "minimal" mode where you could still access callers, data flows, and dependencies with no AI text, would it be helpful for your reviews?
Not parent, but in my opinion the answer here is yes. I agree that there is a real need here and a potentially solid value proposition (which is not the case with a lot of vscode-fork+LLM-based starups) but the whole point should be to combat the verbosity and featurelessness of LLM-generated code and text. Using an LLM on the backend to discover meaningful connections in the codebase may sometimes be the right call but the output of that analysis should be some simple visual indication of control flow or dependency like you mention. At a first look the output in the editor looks more like an expansion rather than a distillation.
Unrelated, but I don't know why I expected the website and editor theme to be hay-yellow and or hay-yellow and black instead of the classic purple on black :)
Thanks for the opinion! That makes a lot of sense and I like the concept of being an extension of a user's own analysis vs hosing them with information.
Yeah originally I thought of using yellow/brown or yellow/black but for some reason I didn't like the color. Plenty of time to go back though!
honestly i feel the same way and i can't quite put into words why. I guess if I had to -- I think it's because I know not all AI generated stuff is equally created and that some people are terrible at prompting/or don't even proofread the stuff that's outputted, so I have this internal barometer that screams "you're likely wasting your time reading this" and so I just learned to avoid it entirely. Which is sad, because clearly now a ton of stuff is AI generated, so I barely read anything, _especially_ if I see any signals like "it's not just this, it's that"
Products like these make me realize we're solving for the wrong problems with a lot of these AI solutions. I don't want you to take this as a hit to you or your product, I actually think it's extremely cool and will likely find a use. But from my perspective if this is a product you think you need, then you likely have a bigger organizational issue, as PRs are probably the last thing that I would want an AI 'intern' to organize for me.
> you likely have a bigger organizational issue
Could you expound on this? In my experience as a software engineer, a pull request could fall into one of two buckets (assuming it's not trivial):
1. The PR is not organized by the author so it's skimmed and not fully understood because it's so hard to follow along
2. The PR author puts a lot of time into organizing the pull request (crafting each commit, trying to build a narrative, etc.) and the review is thorough, but still not easy
I think organization helps the 1st case and obviates the need for the author to spend so much time crafting the PR in the 2nd case (and eliminates messy updates that need to be carefully slotted in).
Curious to hear how y'all handle pull requests!
This is where I feel like we've solved a third-order problem. If you're sorting all PRs into those two buckets then you should probably take a step back and redefine what a PR is for your organization, as both 1 and 2 make the assumption that the PR is too big to review in a single sit down or that the author didn't put in enough effort to craft their PR. Both of these should just be rejected outright in favor of doing things in a smaller more manageable way, instead of having an AI sort through something that a human should have started with. Obviously this is more of an ideal situation and a lot of companies don't work on the ideal which is why I think your product will find good use because companies don't want to invest in slowing down, only going faster.
> you should probably take a step back and redefine what a PR is for your organization
I agree with this wholeheartedly if you are in a role that allows you to redefine what a PR is. In almost every organization that I've worked for, the PR is defined several levels above my pay grade and suggesting changes/updates/etc is usually seen as complaining.
Interesting. At my previous company there was a debate about smaller PRs vs bigger PRs and the end conclusion was that there are tradeoffs in being able to deal with 2-5 bite-sized PRs vs one large PR. The largest one being that it's hard to grasp the totality of the pull request and how the different PRs work together.
> companies don't want to invest in slowing down, only going faster.
I do think this is the way things are going to go moving forward, for better or for worse!
My solution is to organize my PRs as a sequence of commits that explain what they do, and then a PR description that gives an overview and motivates the changes. I've gotten really positive feedback on this, and it dramatically speeds up reviews of my code. Overall less work for the team. (And it often helps me find problems before I even submit the PR.)
As for other people's PRs? If they don't give a good summary, I ask them to write one.
> If they don't give a good summary, I ask them to write one.
Exactly; if people can't be bothered to describe (and justify) their work, or if they outsource it to AI that creates something overly wordy and possibly wrong, why should I be bothered to review it?
How much time does it take you to craft the PR compared to writing the code? Just curious if crafting the PR becomes significantly easier with practice.
Yeah I did this too as an engineer!
I think this is a valid part of the "crafting PR" skill that's under appreciated, and part of the goal of Haystack here is to make that part of PR craft effortless.
The effort is part of what makes the outcome better, though.
Very glad you're working on this! Here is my wish list as a code reviewer:
1. Allow me to step through the code execution paths that have been modified in the pull request, based on the tests that have been modified.
2. Allow me to see the data being handled in variables as I look through the code.
3. Allow me to see code coverage of each part of the code.
4. Show me the full file as I am navigating through the program execution so that I can feel the level of abstraction and notice nearby repetition or code that would benefit from being cleaned up.
Full article: https://dtrejo.com/code-reviews-sad
> 1. Allow me to step through the code execution paths that have been modified in the pull request, based on the tests that have been modified.
Not sure if I fully grasp this! We tried to kind of do this in previous iterations (show call graphs all at once) and it gets messy very fast. Could you elaborate on this point in particular?
Sure - imagine my PR adds one new test which test one new function.
Starting from the test, allow me to step through the program execution, just like a debugger, to observe variables, surrounding code, and the complete file.
If you read only the covered lines of code in a linear way, you'd miss the refactoring opportunities because you aren't looking at the rest of the file.
Did not load, sad.
Failed to load resource: net::ERR_BLOCKED_BY_CLIENT ^ I'm not exactly sure what this is about. I think it is https://static.cloudflareinsights.com/beacon.min.js/vcd15cbe... which I would imagine is probably not necessary.
Uncaught TypeError: Cannot convert undefined or null to object at Object.keys (<anonymous>) at review/?pr_identifier=xxx/xxx/1974:43:12
These urls seem to be kind of revealing.
Trying to fix/find this! Could you email me the repo details? Very sorry for the error here.
In terms of auth: you should get an "unauthenticated" if you're looking at a repo without authentication (or a non-existent repo).
This nails a real problem. Non-trivial PRs need two passes: first grok the entrypoints and touched files to grasp the conceptual change and review order, then dive into each block of changes with context.
It could be just fatigue over the technology itself, but “like you wrote them yourself” sounds too much like a dog whistle for a user base of programmers working with AI generated PRs. Which means that this would be AI iterating over AI.
Feedback: Try speeding up your demo animations and resize the mouse to its regular size. My estimate is that if the marketing copy explains what a thing is, what it does and why it’s useful then all a visitor wants to see in an image is things go pop, boom and whoosh.
I don't think this isa fair critique at all.
Code reviews have always been about primarily reviewing others people code.
Abd knowing your won code better than others people code is a real thing?
Thanks for the feedback!
Not sure if I fully grasp what you mean by dog whistling, but at the end of the day, like another commenter said, Haystack is also pretty helpful for when you're done experimenting with a piece of work and need to see what an AI has generated.
I love this idea, trying it out now. There is QUITE a delay doing the analysis, which is reasonable, so I assume as a productionized (non-demo) release this will be async?
There's a GitHub app that you can install on your repo.
If you install and subscribe to the product, we create a link for you every time you make a pull request. We're working (literally right now!) on making it create a link every time you're assigned a review as well.
We'll also speed up the time in the future (it's pretty slow)!
I would really want to use this, maybe about once a week, for major PRs. I find it absurd that we all get AI help writing large features but very little help when doing the approx same job in reviewing that code. I actually would even read my own PRs with it, as my workflow with AI is to prompt it to acheive building some feature/goal, then only review the code once things work (this is an oversimplification).
Yes! I personally iterate without understanding the code too much (since it'll change), and fully assess what Claude Code (in my case) has done after I finished a piece of work.
I think tools like this are useful, but they can never replace the quality of the narrative that someone who actually wrote the code can come up with.
There's just so much contextual data outside of the code itself that you miss out on. This looks like an improvement over Github Co-Pilot generated summaries, but that's not hard.
I immediately thought this was an update by Deepset and their Haystack framework. https://haystack.deepset.ai/
Just FYI.
This is a cool problem to solve! especially as more and more code is being generated with AI. But are you pivoting away entirely from code visualization? Curious to hear why it didn't work out.
We may go back to code visualization for editors, but after testing it with a lot of engineers, we found that it was too confusing for code reviews. The main issue was just how alien the interface is!
This is an amazing idea and I would like to try it.
I hope you (eventually) ship something for AR to visualize software components interacting in 3D space
If AI writes all of the code, we will need to max out humans’ ability to do architecture and systems design.
>If AI writes all of the code, we will need to max out humans’ ability to do architecture and systems design.
Strongly agree with this. There are demos that you're able to try, by the way!
I have vague recollections of an infinite canvas IDE with the same name.
Are you the same folks that worked on that?
Yes! We haven't wholly given up on that idea by the way ;)
I love the AI of adding context and logical ordering to PRs! Really cool concept
Any ideas what pricing will look like?
What is your privacy policy around AI?
Any plans for a locally-runnable version of this?
1. Pricing would be $20 per person and we'd spin up an analysis for every PR you create/are assigned to review 2. We don't train or retain anything related to your codebase. We do send all the diffs to Open AI and/or Anthropic (and we have agentic grepping so it can see other parts of the codebase as well) 3. Do you mean the ability to run this on your own servers (or even on your own computer with local LLMs)? We do have plans for the former, but I don't know how big the lift on the latter will be so I'm not sure!
> We don't train or retain anything related to your codebase.
It would be good to have this mentioned on the website somewhere, as part of a privacy policy. Right now I can't find any details on the site, which is preventing me from trying this out with a production repository.
Will add this then!
This is cool. What are you using to trace data flow?
We use a mixture of LLMs (e.g. to identify the important variables to trace), language servers (to understand where a variable is originally defined and where it's used).
Pretty cool. Not from a big firm, but I really like the idea of adding metadata to PRs! One counterintuitive thing about navigation is that I keep hitting the Back button, hoping to go back to `Pull request summary`, which feels like the main navigation page, but nothing happens. You already implement navigation back and forward in history, so why not do it in the browser router too?
Sorry to clarify: you hit "back" after traversing from the summary but it doesn't go back? If so, that's a bug!
Or do you mean that doing the browser navigation of "back" should bring you to the summary (initial page)?
Sorry for the confusion! Missing browser navigation is the problem. The virtual back buttons you put in the top left are working as I'd expect browser nav to do. I keep trying to go back, it would feel so natural.
Ahah! Thanks for the feedback.