next level nerdingYour Average Tech Bro's coding setup
Your Average Tech Bro's coding setup
Thomas builds Yorby, an app making about $10K a month, with one co-founder and no other engineers. The setup: agents in the cloud that test their own work, and one bot that owns each feature. Thomas writes a careful spec, looks at the proof, and says merge.
Jump to the workflow map, the skills, or try it yourself.
Original sources
The video
21 minutes, posted Sept 24, 2026. Every timestamp on this page opens it at that moment. 5:36 to 7:18 is a sponsor read.
- 0:00 The updated stack, and what got dropped
- 1:51 Why both Codex and Cursor
- 2:40 Codex for local work
- 4:59 Cursor cloud agents
- 5:36 Sponsor read (until 7:18)
- 7:18 pstack and planning with several models
- 9:25 How the agents test their own work
- 12:36 Bots set up like an engineering team
- 17:44 Customer support and error fixing, automated
Eight signals
The whole video in eight boxes. Each links to the part of this page that explains it.
- 012:40
Two tools, split by where the work runs
Codex on the laptop for front-end work you have to look at. Cursor in the cloud for everything else. $200 a month each.
How - 0210:50
Verification skills are the investment
Written instructions that tell an agent how to open the app, use a feature and check it works. Thomas calls building these the most important thing an AI-first team can do.
How - 038:20
Plan with many models, build with cheap ones
Several models each draft a plan. A stronger judge merges the best parts into one. Cheaper models then write the code.
How - 0411:18
The agent proves it, you don't click
A cloud agent has its own desktop. It runs the app there and sends back screenshots or a recording. No more testing by hand.
How - 0513:04
One bot owns one feature
Each part of the app has a Grok Bot that explored the code, used the feature and keeps notes on how it works. You message that bot like a teammate.
How - 0614:54
Every merge teaches the owner
After each PR the owner bot does two things: update its notes on the feature, and add or fix the verification skills for it.
How - 0716:40
The engineering moves into the prompt
A feature prompt takes 10 to 15 minutes to write: what to change, which parts to touch, what must pass, what must not break.
How - 0819:20
Auto-merge only for a real emergency
The error bot runs all day and night but opens PRs for review. It may merge on its own only when a bug is hitting most users.
How
Words on this page
Thirteen terms, in plain words. Skip this if you already run coding agents and review PRs.
- Agent
- An AI that does coding work on its own: reads the code, makes changes, runs the app. On this page Codex and Cursor run them.
- Cloud agent
- An agent that runs on a machine in the cloud, not on your laptop. Cursor gives each one a virtual machine with a full desktop, so it can open a browser and use your app.
- Worktree
- A separate working copy of the same project, so several agents can change code at once. Codex stays fast with many open; Cursor slowed down.
- Front end
- The part of the app people see and click. Thomas does small front-end changes on his laptop, where he can look at each one quickly.
- Model
- The AI that does the thinking inside an agent. Cursor offers many, some stronger and pricier, some cheaper.
- Judge model
- A stronger model that reads plans from several other models and writes one final plan from the best parts.
- Tokens
- What models count and charge for. More models working on the same task means more tokens spent.
- Spec
- The written task you hand the agent: what to change, which parts to touch, what must pass, what must not break. Thomas spends 10 to 15 minutes on one.
- Surface
- One part of the app, like a screen or a feature. A spec lists the surfaces in play, and verification skills are written per surface.
- PR (pull request)
- A proposed code change waiting for approval. Here each PR arrives with screenshots or a recording as proof.
- Merge
- Accepting a PR so the change becomes part of the real app. After each merge the owner bot updates its notes and skills.
- Knowledge base
- The notes an owner bot keeps on how its feature works. It builds them by reading the code and using the feature, and updates them after every merge.
- Escalate
- Pass a case up to a person. The support bot escalates any email its instructions don't cover to the two founders.
What's in, what's out
Two paid plans at $200 a month each, out of pocket. Thomas says up front this is built for speed, not for a tight budget. 2:22
Local agent$200 / month
Codex desktop app
For work on the laptop, mostly small front-end changes. It stays fast with many worktrees open, where Cursor slowed down. It has a device toolbar that changes the preview size while the chat stays visible, and a quick tool for marking up the page.
Where most work runs$200 / month
Cursor cloud agents
Coding agents that run in the cloud. Each gets a full virtual desktop, so it can open a browser and use the app it just changed. You can open that desktop and look.
The skills the agents followFree, open source
pstack
Lauren Tan's bundle of skills for Cursor. Two parts matter most here: planning with several models, and making the agent prove its work with verification skills.
The front door
Grok Bot
Made by the same company as Cursor, so a bot can start a Cursor cloud agent for you. Thomas keeps one bot per feature and talks to them in chat.
Error feed
PostHog
Captures every error from the live app. The error bot reads this feed.
Customer email
A support inbox
Plain email, watched by the support bot. It replaced Intercom.
The workflow map
Eight steps in a loop, plus three boxes beside it. Darker boxes are the two steps that need you. Click any box to read it, or press Start and follow one bug from report to merge.
Cloud lane · most of the work · Cursor cloud agents + pstack + Grok Bot
Beside the loop
Follow one bug through the loop
Eight beats, using the bug Thomas fixes on camera: a model picker that disappears when the chat window is minimized. Count how many beats need you.
Cloud lane · step 1 of 8 · You
Watch at 16:40Write the spec
You write the task like a ticket for a junior engineer. What to change. Which parts of the app to touch. The scenarios that must pass. The scenarios that must not break.
Why: This is where the engineering now happens. A detailed prompt forces you to think the change through before any code exists. With a good one, the rest mostly runs on its own.
What you see
A long chat message. Nothing else yet.
Watch out
A one-line prompt gets a guess. The time you save typing you lose reviewing.
The skills
What a verification skill is
A recipe for using one part of your app: how to get in, where to click, what a good result looks like. An agent follows it to test a change the way a person would.
Why it matters most
Thomas rates it as “probably the most important thing you can do” for a team built around agents. It is the reason the proof can be trusted, and the reason the error bot is allowed to merge at all. 10:50
How they get written
Not by hand. The agent tests a feature, then writes down what it did so the next agent can repeat it. Owner bots keep them current after every merge.
Writes a verify skill for your project when it has no scripted way to prove the app works. Includes a map of the app's features.
When: Once per app or surface, the first time an agent has to test it.
Fixes a verify skill that has drifted from the app. Reads the code, does one live pass, and proposes only corrections it has proven.
When: After merges that change how a feature works.
Runs several attempts at the same thing in parallel, then takes the best parts of each. Thomas doesn't name it; it's the pstack skill that matches the planning step in the video.
When: Features and anything with more than one sensible approach.
How to open and test the Yorby web app. The folder holds a SKILL.md plus artifacts, bin and features folders.
When: Any change a user would see in the browser.
How to check the video rendering that runs on Remotion Lambda.
When: Changes to how videos get made.
How to check email sending through Resend.
When: Changes to emails the app sends.
How to check connecting a TikTok account.
When: Changes to the TikTok link-up.
The four Yorby skills are named on screen but their contents aren't shown, so the descriptions go by the names. For what a verification skill looks like inside, see the pstack page. pstack has about twenty more skills; Thomas plans a separate video on it.
The bot team
Each bot is a Grok Bot with one job. Thomas compares the feature owners to staff engineers or tech leads: each one knows its part of the app and gets the work for it. 13:04
Feature owners · you message them
- UGC StudioThe AI content studio, where a startup makes marketing videos in set formats.
- Account SpyWatches competitors' social accounts and alerts you when their posts take off.
- AnalyticsThe analytics feature.
- Content CoachThe AI script writer.
- Content CalendarMakes content for you in the background. Still being built.
Always on · they message you
- Yorby Error FixerEvery error from PostHog. Routes to an owner, or fixes what nobody owns.
- Customer SupportThe support inbox. Handles known cases, escalates new ones.
How an owner is made
The bot explores the code for its feature, uses the feature itself, and builds a knowledge base from both. From then on, work on that feature goes through that bot.
Not tried yet
For a feature that crosses two owners, put both bots in a group chat and have them plan together and check each other. Thomas describes the idea and says it hasn't been done yet. 15:37
The spec
Thomas spends 10 to 15 minutes on one feature prompt, and treats that as the real engineering. The video names what goes in but never shows a full prompt, so the template below is FakeNerd's. 16:40
- 01What to change
The behavior you want, in plain words. What a user does and what they should see.
- 02Surfaces to touch
Which parts of the app are in play, so the agent doesn't wander into others.
- 03Scenarios that must pass
The exact cases the agent has to run and show you proof of.
- 04Scenarios that must not break
Nearby behavior to re-check, so the fix doesn't cause a new bug.
- 05Context and references
Links, screenshots, related code or past PRs the agent should read first.
FakeNerd starter · the spec
TASK [One paragraph: what should change, from the user's point of view.] SURFACES TO TOUCH - [feature or screen] - [feature or screen] Leave everything else alone. MUST PASS (run each one in the real app and send proof) 1. [Scenario: steps, expected result] 2. [Scenario: steps, expected result] MUST NOT BREAK (re-check these after the change) 1. [Nearby behavior] 2. [Nearby behavior] CONTEXT - [Link, screenshot, related PR or file] DONE MEANS A PR, with screenshots or a recording of every MUST PASS scenario working. Use the verification skill for this surface. If none exists, create one first.
FakeNerd starter · after the PR is approved
Merge this PR. Then, as the owner of this feature, do two things: 1. Read the changes and update your knowledge base on how this feature works. 2. Create or update the verification skills for this surface, so the next agent that works here has a written way to use and test it.
This is the standing instruction from step 8, in our words. Thomas gives it to the owner bot on every merge. 14:54
Two bots that never clock out
These run the company side: errors and customer email. Neither waits for a prompt. Each has a short list of questions it answers the same way every time.
Error bot 18:40
1An error arrives from PostHog. Does it belong to one feature?
Yes
Package the details and hand it to that feature's owner bot to fix.
No
The error bot fixes it itself. It is also the catch-all bug fixer.
2A fix is ready. Is the bug urgent and hitting most users?
Yes
Merge without waiting for a review. This is rare by design.
No
Open a PR and wait for Thomas.
The bar for merging alone is set very high on purpose: urgent, and affecting all or most users. Thomas ties that trust directly to the verification skills.
Support bot 17:58
1A support email arrives. Is this case in the instructions?
Yes
The bot handles it.
No
It brings in Thomas and the co-founder, who solve it by hand.
2The founders solved a new case. What then?
Then
They add the case and the answer to the bot's instructions. Next time it's handled.
Same shape as the owner bots: every case a person solves becomes a written rule the bot follows next time.
Try it, smallest version first
Six steps, in the order that costs least. You can stop after step 4 and already have agents that test their own work.
- 01
Pick one tool for cloud agents
Thomas uses Cursor. What you need is an agent that runs on its own machine with a real desktop, so it can open your app.
cursor.com - 02
Install pstack
It gives the agent the planning and verification habits described here. Setup takes about ten minutes.
pstack setup on FakeNerd - 03
Make your first verification skill
Run /create-verification-skill on the part of your app you change most. Expect this to be slow. The agent is learning how to use your app.
What a verification skill looks like - 04
Run one task with a full spec
Use the template above. Don't accept the result without screenshots or a recording of each scenario.
- 05
Add one owner bot
Start with your busiest feature. Have the bot explore the code and use the feature, and keep notes. Thomas uses Grok Bot; any agent that keeps notes between chats and can start a cloud agent will do.
- 06
Add the always-on bots last
Errors first, then support. Keep auto-merge off until your verification skills have earned it.
FakeNerd's take
The idea worth stealing is step 8
Most setups stop at merge. Making the agent update its notes and its test recipe after every merge is what keeps the whole thing from rotting. It costs one extra sentence in a prompt.
It's $400 a month before you start
Two top-tier plans, and Thomas says plainly this isn't a budget setup. The cheap version: one cloud agent tool, pstack, and one verification skill. That gets you steps 1 and 3 to 7.
Trust the proof, check the proof
Screenshots show the case the agent chose to test. They don't show the case it didn't think of. Your must-pass and must-not-break lists are what make the proof worth something.
Mind who's selling what
Grok Bot sponsored an earlier video on the channel, which Thomas says on camera. Revenue and speed numbers are self-reported. “30+ PRs a day” is in the title and never said out loud.
Other people's versions of the same idea on this site: Ras Mic's software factory (proof on every PR, an outside reviewer scoring it) and Warp's software factory. The skills come from pstack.
Attribution
- My New AI Coding Setup That Helps Me Ship 30+ PRs A Day, Your Average Tech Bro (Sept 24, 2026). The tools, the bot team, the two always-on bots and the reasoning are from this video, in our words. The one quote is from its captions. Skill names, bot names and the bug are read from the screen. Sponsored by Fish Audio, which isn't part of the workflow.
- pstack by Lauren Tan (@poteto), MIT. What the three pstack skills do is from its README (v0.15.5), summarized.
- youraveragetechbro.com and yorby.ai. Revenue, prices and the PR count are Thomas's own; nobody has audited them.
- The map, the eight-beat walkthrough, both starter prompts, the setup order and the takes are FakeNerd's. Not sponsored by anyone on this page.