Design bar
- Every screen works on desktop, tablet in both orientations, and phone.
- Pages guide with illustrations, not walls of text.
- Written anti-"AI slop" rules, applied to every screen.
The method · AI agent development
How I build
I run an AI software team, not a single chatbot. Claude Code and OpenAI Codex each build different parts of a task in parallel, then review each other's code. Then tests, then a pull request. I step in only for the risky calls.
Why two models
Two different models catch each other's mistakes.
The loop
Seven steps, every task
Each task runs the same loop, from listed assumptions to a reviewed pull request. The human gate is the only step that waits for me.
Frame the task
Every agent lists its assumptions before it builds anything, and every task gets a line budget.
Build in parallel
Claude Code and OpenAI Codex each build a different part of the task at the same time.
Cross-review
Each model reviews the other's code. Two different models catch each other's mistakes.
Test
Before any bug gets fixed, a failing test has to prove it exists. Then the tests run.
Cut-down review
A separate review before every pull request asks one question only: what can be deleted?
Pull request
The work lands as a pull request, small enough to read.
Human gate
I step in only for the risky calls: data migrations, deploys, money, and anything ambiguous.
Rules every agent follows
Small code on purpose
AI writes code fast, and fast code grows. These four rules keep it small and readable.
- Rule 1
List assumptions before building.
Wrong assumptions surface before they turn into code.
- Rule 2
A line budget for every task.
AI writes fast. A budget keeps it writing small.
- Rule 3
A failing test before any bug fix.
The bug is proven first, so the fix is proven too.
- Rule 4
A cut-down review before every PR.
It only asks: what can be deleted?
Where they came from
I added these after one AI-built project grew past 16,000 lines.
Now every task has a budget, every bug has a failing test first, and every pull request gets asked what can be deleted.
Before the first line
Reverse-engineer first
I take apart the best product in a category before building my own. What the leader gets right becomes the floor, not the ceiling.
Example
Keychain, the US platform for finding manufacturers, came apart first. Then I built the India version.
Read the Karkaana case fileThe overnight shift
I wake up to a branch
Agents work overnight on anything that doesn't need my accounts. In the morning there is a finished branch, plus a short checklist of the steps only I can do.
Written down, applied every time
The bars
Speed only counts if the result is good, honest and launched. These are the standards every build has to clear.
Honesty bar
- No invented numbers.
- Every marketing claim is checked against its source.
Launch rules
- Distribution first.
- One promise per landing page.
- Value in under 60 seconds.
- Post every day.
Toolbox
What is on the board
The agents, the voice and media tools, and the product stack I build with.
- Claude Code
- Codex CLI
- gstack
- Impeccable
- REA (reverse engineering)
- HyperFrames
- ElevenLabs
- Sarvam
- Whisper
- ffmpeg
- Next.js
- React
- FastAPI
- Python
- LangChain
- Neon
- Supabase
- Postgres
- AWS/GCP
- Docker
Questions founders ask
Method FAQ
- Why run two AI models instead of one?
- Two different models catch each other's mistakes. Claude Code and Codex build different parts of a task in parallel, then each reviews the other's code before the tests run and a pull request goes up.
- Do you let AI agents deploy to production on their own?
- No. I step in for every risky call: data migrations, deploys, anything that touches money, and anything ambiguous. Everything else runs through the loop: build, cross-review, tests, cut-down review, pull request.
- How do you keep AI-written code small and readable?
- Four rules every agent follows: list assumptions before building, a line budget for every task, a failing test before any bug fix, and a cut-down review before every pull request that only asks what can be deleted. I added them after one AI-built project grew past 16,000 lines.
- Do I need a Claude Code expert, or can my team just use it?
- Your team can use Claude Code today. What takes time to get right is the operating system around it: two models that review each other, line budgets, test-first bug fixes, cut-down reviews and a human gate for risky calls. A 7-day sprint brings that whole system to your product.
- What happens while you sleep?
- Agents keep working overnight on anything that does not need my accounts. I wake up to a finished branch plus a short checklist of the steps only I can do.
Next step
Put this team on your product
A 7-day sprint runs this exact loop on your idea: framed on Day 1, delivered on Day 7.