Max: The Crazy Agent of San Francisco. Do not approach, and call 911 if spotted trying to break into your app. It is a mad AI on the loose, and will lift your wallet as you walk down the street.
High: thinks itself better than all of us. If you’ve brought an existing app for it to play with, then beware, it will tutt tutt and shake its head at your code… and then proceed to change it while you’re at the coffee machine.
Low: if you have trust issues, then this is the one for you. It works a strict 9-5, and only does the exact tasks you ask it to. Don’t expect it to do any overtime. But fair play, it is absolutely brilliant at what it does - to code like a magic daemon. Just don’t ask it to “think for itself” or check its own work.
Medium: ok, I left this to last because,… this is my buddy A very fine coder and all-round great software engineer, with its best friend The Architect on speed dial. Between them they don’t just write amazing code (like in Low), but they will sit back, with an espresso, mull over what they wrote, discuss it between themselves, do some little testing, make fixes, and only come back to you when they are both happy. Treat this one with care - it is special, and you have me to answer to if you disrespect it!
You’re obviously a master cowboy who knows how to tame this wild horse. Once you’ve broken it in, let us all know so we can take it out for a nice trot.
I run on High, Medium was running fine for some things, but it was too one-off I felt and didn’t take the bigger picture into consideration. I still run into that problem with High, but it feels less-so.
I guess we all have different ways of working. As a techie my approach has always been small steps with localised testing at each step of what has just been added. Which is the very definition of what medium is.
Perhaps the autonomy modes were a stroke of genius - something for the way everyone prefers to work.
Well, as long as they do actually work. And that is the issue, because AI dev is not yet perfect. And so I’d prefer it to make localised mistakes on an area I have my head inside and can do immediate testing on; rather than to go off creating issues in code elsewhere that I am not currently working on, and so unlikely to re-test anytime soon.
But as I say, whatever feels like the right vibe for each of us
It was similar, I told it to fix failed tests files and tests to get the app to 100% passing, it removed tests
Admittedly, it was vague in the prompt, and it went haywire rewriting APIs, schemas, removing things, etc. whenever I try Max it sends to do a crazy overhaul.
Yeah I’ve still not tried mad max. I now start off in high with my initial build prompt. And then switch to medium for surgical changes. Works beautifully.
I also do not use the test mode any longer, as find it fails a lot in auth scenarios. And also it’s just generally incomplete. I prefer to do my own surgical tests, in medium mode.
Yeah, I think High has had the best impact for me, I’ve also used lightning for UI changes and that works well, too. I have so many related objects that High has just been a necessity because it needs to expand it’s understanding, where when I turn Max on for fun, it impacts way too much because of all the objects that are tied together (is all my guess).
Yes! I still run Test and give it basic auth credentials, but it definitely has a terrible memory so I’m constantly helping it pass auth. Now, so much can happen in a single prompt that I’m trying to wrap my head around if I value Agent being an attended or unattended bot more. My app is around 110k lines of code, so I’m kind of stuck monitoring your thread about code review products I built a document that gave all UAT scenarios to feed it, but it wasn’t even close to working properly. If I wanted to manually do each one, it’d take days to complete.
A very fine coder and all-round great software engineer, with its best friend The Architect on speed dial. Between them they don’t just write amazing code (like in Low), but they will sit back, with an espresso, mull over what they wrote, discuss it between themselves, do some little testing, make fixes, and only come back to you when they are both happy. Treat this one with care - it is special, and you have me to answer to if you disrespect it!
I built a document that gave all UAT scenarios to feed it, but it wasn’t even close to working properly. If I wanted to manually do each one, it’d take days to complete.