Convex projects just got a BIG AI upgrade!

So, you want to level up your Vibe

coding sessions, but can't be bothered

to go out and find out whatever the

current best practices are when working

with Convex projects. Well,

>> good news everyone.

>> Those Convex best practices now

automatically come to you when you run

Convex Dev. That's right. As of Convex

1.34, if you have a new ComX project and

you run Convex Dev to set up your new

deployment, you're now given this new

prompt to install the AI files. These AI

files are both the guidelines which we

have painstakingly refined uh from our

hard one bottles in the convex evals

repo and our newly minted agent skills.

Then once you have the AI files

installed, it'll go ahead and prompt you

whenever there's a new version of those

files so you always have the latest

greatest AI enhancements. But you want

more details or just have no idea what

I'm talking about? then stick around as

I want you to grab yourself a lovely cup

of tea, drop me a like and sub, then

we're gonna get into it.

So, for the longest time now, we've been

automatically shipping our cursor rules

files into your Convex projects when we

detect you're running inside a cursor.

Now, this may have been all well and

good back in the dark ages of 2025, but

you might have noticed things have moved

on a little bit since then. Now, we have

many different kinds of agents running a

host of different models that could all

do with the guidance provided by our AI

guidelines. If you aren't familiar with

our AI guidelines, then I've done a

number of videos on this already, but

the TLDDW is that models are really good

at convex, but not absolutely perfect.

So, what we did is we designed a series

of test cases called evals that poke the

model to see what it does and doesn't

know about convex. Then if the model

fails at one of these tests, we add a

guideline to plug in this gap in the

model's knowledge. Now, this has been

really successful because not only do we

get a really useful set of guidelines

that we can pass to a model to improve

it, it also allows us to generate this

LLM leaderboard which we can use to see

which models currently do best and worst

[music] at convex both with and without

the guidelines. And also the big AI labs

absolutely love eval sets like this as

it allows them to continually refine

their models so that in theory over time

we might just be able to remove these

guidelines because the models already

score perfect on convex. Anyways, back

to the point. So we were shipping these

cursor rules into your projects

automatically which are basically just

the guidelines with a bunch of cursor

specific fluff at the top. But these

would obviously only apply to tools that

would pick up these cursor rules files.

So tools for example like claude code

wouldn't pick these up. So now we

changed it. So when you run

npxconvexdev,

you select yes from the menu. We add the

rules as a markdown document to

underscore generated/ai/guidelines.mmd.

But the question is how do models know

to look there and read that? Well,

that's why we also inject this little

snippet into your agents.mmd file and

your claw.md file which basically covers

the vast majority of agents and it will

encourage the models to always read

those guidelines first when working on

convex code. So I guess naturally at

this point you probably have lots of

questions. Why do we do it this way? Why

don't we just simply dump the entire

guidelines into your agents.mmd file?

Why don't we do an index of the convex

docs like Versel does? Why do we put the

generated files in /generated AAI and

why they not get ignored? Well, these

are all really good questions and we

thought long and hard about each of

these, but I think this video would take

forever if I went to details about each

of them. So, drop me a comment down

below if you're interested and I will

happily give the reasoning for them.

Okay, so this feature is called AI files

and not simply AI guidelines. That's

because we also ship our newly minted

agent skills into your project, too. So,

if you aren't familiar with skills,

they're like little bundles of knowledge

that can assist a model with a

particular action you're trying to

undertake. Now, I chose my words

carefully there because, as mentioned in

the philosophy section of the readme, we

want skills to be laser focused on a

particular action you want to take. So

basically skills are not a place for

simply repeating what is already in the

convex docs or filling in gaps in the

model's knowledge. That's what the eval

guidelines are for as previously

previously mentioned. So with that in

[music] mind, let's take a look at what

skills we have in here currently. So

convex setup orth walks you through

choosing an orth provider, then the

steps required to get that particular

orth provider working. Be it convex or

works, clerk or zero. Convex performance

audit is a really handy skill that

encourages the models to go out, gather

a bunch of signals such as the newly

minted NPX convex insights command. Man,

I need to do a video on that one soon, I

think. And the convex doctor. Uh, I

should probably do a video on this one

as well, thinking about it. Before

highlighting potential issues for you,

we've also provided a number of

reference documents that have been

carefully gleamed from our knowledge and

even some recent learnings from Seth's

optimizations of Clawhub. You should go

out and read that article by the way.

It's very interesting. We also added a

skill to help with migrations,

uh, one to help with creating components

and one to help kickstart Convex

projects from scratch. Now, as mentioned

before, these skills are automatically

installed for you when you run npx

convex dev and select yes to AI files or

if you run npx convex AI files [music]

install. More on these various CLI

commands a little bit later. But I

should just note that this AI files

feature just leverages our agent skills

which are in a separate repo. So you

aren't required to use this feature to

be able to use the skills. that you

could just simply copy them into your

directory if you want or you can

leverage the npx skills tool to add them

yourself. And fun fact, we actually use

npx skills ourselves underneath the hood

when we install them into your repo. We

decided to do this so that we could

leverage this tool rather than write our

own because unfortunately installing

agent skills at the moment is a little

bit of an art than a science and it's

constantly changing mainly thanks to the

fact that the various IDs tools or

whatever just can't seem to agree on a

on a correct standard. So for now, we

decided just to let the NPX skills CLI

worry about what where the best place to

install these things given on somebody's

setup, and we'll just carry on doing

what we do instead.

So, one thing I was keen on getting

right from the start was the update

story. These skills are going to change

over time as Convex and AI evolves. So,

I didn't want this to be like a a

oneshot install and then it goes stale.

So after the initial setup of your

project, every subsequent time that you

run npx convex dev, it's going to go out

and check to see whether there's any

updates to the guidelines or the agent

skills. And if there is, it's going to

drop this little warning in your in your

terminal to let you know. Then you can

just run npx convex AI files update to

get the latest versions. We purposely

chose not to automatically update those

files for you for now at least as we

didn't want to create surprise

unexpected changes to your repo when

running npxdev.