Thomas SchlossmacherFeb 5, 2026
Two years of vibe coding into 9 lessons
16views
Two years of vibe coding into 9 lessons
Thomas Schlossmacher
Learn 9 valuable lessons for improving your workflow with AI agents, from organizing projects and using multiple agents to defining agent identities and optimizing code for AI navigation. This video shares practical tips to enhance your development process.
Two years of vibe coding into 9 lessons
Thomas Schlossmacher
Learn 9 valuable lessons for improving your workflow with AI agents, from organizing projects and using multiple agents to defining agent identities and optimizing code for AI navigation. This video shares practical tips to enhance your development process.
Chapters
00:00
Introduction
Introduction to lessons learned in AI development.
00:21
Organize Projects
00:56
Multiple Agents
01:33
Agent Identity
02:10
Isolate Scope
03:16
Spawn Sub-Agents
03:34
Rule of Threes
04:19
Skills and Docs Loop
Transcript
00:00
In the last two years I have really grown from being a idiot to being someone who's a little bit more comfortable working with anything.
00:09
Technical IDEs, libraries, agents, whatever.
00:13
These are nine lessons that I learned and I wanted to share them with you.
00:16
I put them up on my site if you want to read through this article, but this is a quick video.
00:20
First thing, if you have an app and you have a site or you have a few different projects related to one project, put them all in one big folder.
00:29
Tell your agent in the Claude MD file or the agent file.
00:33
There are other projects in this directory that you can reference.
00:37
If you reference those things like, hey, mimic the design system of the website in the app.
00:44
It's much easier because your agent can actually reference real things.
00:49
Hey, reference the authentication we built on the website for the app.
00:52
You can keep everything connected.
00:54
Second thing, use multiple agents at once.
00:57
That might be more relevant this year.
00:59
As you see Codex, Claude code, agent swarms, but have like four panels on your screen, maybe three.
01:06
If you work in anthropic, you can have 10 with unlimited budget, but most people don't.
01:11
If you're using four, you're probably on a max plan.
01:13
If you, if you can't afford a max plan, get two basic plans like Gemini and Claude code and then run them in parallel.
01:20
also with that, use one agent as a planner and the other one as a worker.
01:24
Often people use OpenAI's codecs as the planner and then Claude code as the worker.
01:29
Number three, every agent needs an identity.
01:33
Give them purpose, Tell them, hey, you're the expert at motion design.
01:38
You're world class at this.
01:40
We what are the top things world class motion designers care about?
01:44
Well, it might be fading effects, it might be, you know, sequence of effects, it might be speed, your agent, your language model.
01:52
It will know what the important things that give world class craft to something are.
01:57
And if you give that to the agent as its identity, maybe it's for web design.
02:01
It's going to do a great job with references with anything.
02:04
It's just increasing the level.
02:06
Number four, isolate the scope.
02:07
If you say make me a beautiful website, you're cooked.
02:11
You're not going to get that beautiful website.
02:13
You're going to get a bunch of gradients, really vibe coded slop.
02:17
It's not going to look great and it's going to look really generic.
02:20
What the agent thinks is that it's good.
02:23
What you should do is make a plan.
02:26
Be very specific.
02:27
Again, provide examples, provide an identity.
02:31
show examples of code or link to libraries, components, whatever and tell it, hey, interview me for this particular, component.
02:40
Or you can, you can use the anthropic claude code/feature command to then have a short discussion about what you're building.
02:49
The more time you spend referencing your existing code base, any documentation and planning with the agent on what your expectations are, the less work you're going to do and just waste usage and tokens.
03:00
So if you say make this page beautiful, it's not going to look good.
03:04
But if you say make this component beautiful, you're going to do much better.
03:09
5.
03:09
Spawn sub agents some tasks take a long time.
03:12
Tell it, hey, make the whole plan, but then spawn sub agents and use these identities.
03:17
Those could be saved agent markdown files.
03:19
Maybe a back end and integrations guy and a front end agent.
03:24
number six, the rule of threes.
03:26
I learned this from the every team they have a plugin for compound engineering.
03:31
I think I also saw it referenced by the claudebot openclaw guy.
03:36
what is the rule of threes?
03:37
The rule of threes is telling your agent, hey, you should learn from what mistakes you're making.
03:43
So if you make a mistake three times, I want you to document it, write it in a development log and let's talk about why it happened and then document the fix so your agent automatically knows as it's working, hey, I've run into this problem three times.
03:58
Why don't I go back, write down the problem which is half the solution and then look at the right way to fix this.
04:04
It will make the development process much smoother.
04:08
Next Skills and docs loop.
04:11
if there's available skills from skills shop, you should probably download them, you should use them, you should give them to your agent as Instructions within its AgentMD file and tell it, hey, this is a that is specific for using OAuth on Convex.
04:28
I want you to pay attention to this.
04:30
It will reference that skill when it works with Convex.
04:32
It will reference your devlog and it'll do a good job.
04:35
Skills can make a huge difference.
04:37
They're just fancy prompts for all of your agents outputs.
04:41
Docsloop.
04:42
when you're building something, tell your agent to keep a time based development log what it did, what it changed.
04:49
Add learnings to it so that it has a historical contextual reference of everything that was done and why it's important.
04:57
Why is that useful?
04:58
Well, your agent loses its state when it goes on to a new task.
05:02
If it has a development log, it can see what happened and why it's a log.
05:07
The other thing is kits.
05:09
So that's one thing, the development log.
05:10
The second thing is you should create a map at the top of each internal directory.
05:16
Maybe you have all of your components in one file, all of your very specific app, functions in another.
05:22
Maybe you're breaking up something.
05:24
Have the agent organize big mono files into small very well put well structured files that are properly commented and put a map at the top.
05:35
Or have it create an MD file which is a map for each file and those all lead up to the AgentMD file which has the full map of your project that it auto updates.
05:44
I'm yapping but that is super useful because if you have everything well structured and it can easily be found, things get solved faster and easier.
05:56
number eight Optimize for agents.
05:58
Often you're going to need to refactor your code base.
06:01
Use a agent to refactor files into organized directories with a map but tell it to specifically comment into the code base in a way that is optimized for agents.
06:15
So every couple of commits or every day at the end of the day or at the end of a session, I do it quite often frankly I will tell my agent hey, we did a ton of work.
06:24
What are the learnings by the way?
06:26
Refactor this folder and I need you to optimize this setup for an agent to navigate and work in.
06:35
So instead of providing comments that are really good for a human to read, it actually creates a comment like structure that is very good for an agent to read.
06:43
It's making the environment easier for itself to work.
06:46
The last thing, when an LLM was trained, the knowledge cutoff, the point where they stopped putting in new information is very important because new code bases, new libraries, kits, whatever you're using.
07:02
New documentation in general will not be accessed or not have been trained on.
07:07
So what you want to do is you want to use popular libraries or tools.
07:14
I use a fully kind of integrated agent stack.
07:18
for example convex is in my files with my agent has access to it.
07:23
I use next js, I use resend everything kind of within the vercel world.
07:28
Why super easy.
07:29
It's well documented, it's good for agents.
07:32
So if you're doing obscure work or you're working at the state of the art or you have to build something novel, it's likely it's not documented and your agent won't be able to do as good of a job had it had something that was useful.
07:46
Anyways, I've yapped through this.
07:48
Thank you for your attention.
07:49
If you like agents, if you like coding, if you like AI stuff, leave a comment.
07:54
I'd love to stay engaged with you.
07:56
And if you want, you can look through this entire article.
08:00
I kept it short enough that you can go through in one pass.
08:04
two.
Comments
Comments
No comments yet