RL-AI / AI-native engineering

How I work with AI agents

Agents write a lot of my code. Specs, hooks, tests and reviews decide what ships. Everything below is real and running in this site's repository.

01 / AI-native engineering

Every change goes spec first

1BrainstormIntent and constraints agreed first
2SpecWritten, reviewed, committed
3PlanBite-sized tasks with tests
4Agents buildTest first, every step verified
5ReviewA fresh reviewer agent, then me
6Pull requestCI plus a Claude review on GitHub

This site went through exactly this flow: a design spec, a design-system spec, a pages spec and an implementation plan per phase, all committed next to the code.

02 / AI-native engineering

Unsafe edits are blocked, not discussed

Every edit is formatted and linted by a hook the moment it lands.
Stopping is blocked until typecheck and tests pass.
A guard rejects skipped git hooks, lockfile edits and unapproved config changes.
Architecture tests fail the build when a layer reaches where it shouldn't.

AGENTS.md · Guardrails

## Guardrails

- Change the shared config at the root. Package `tsconfig.json` files only `extends` the base and set `include`; `apps/*` add only what their framework needs (ADRs 0002, 0004).
- Add a tool, a config file or an `apps/*` entry when real code needs it, and record why in a new ADR.
- Secrets live in `.env` files, which are gitignored. Commit a `.env.example` with names only.

03 / AI-native engineering

The architecture, as playbooks

Skills

/new-module
/new-feature
/new-package
/ui-component
/reskin
/verify
/ship

Three pairs of eyes

A reviewer agent checks every diff against the repo's rules.
CI runs a Claude review on every pull request.
I approve every merge.

Real state, not guesses

Agents read the running app, the component kit, the database and the deploy. Production data is read-only, behind an on-demand proxy.

next-devtools
storybook
shadcn
postgres
postgres-prod (read-only)
railway
better-auth-docs

04 / AI-native engineering

The same discipline beyond this repo

Tennis Expert
A nightly agent reads production
A headless Claude Code run reviews the day's production logs and recent commits, and a security-review agent checks every smart-contract change. 522 commits are co-authored with Claude.
AI recruiting platform · under NDA
Changes gated by a benchmark
A custom skill re-runs a fixed benchmark with a holdout set before any engine change ships. Architect, executor and reviewer agents are split by role.
MicroTrading
Pre-registered, then measured
A frozen spec written with Claude, experiments registered before they ran, and a language-model analyst shipped switched off, behind a typed guardrail.

Want this way of working on your team?

Get in touch See the work

## Guardrails - Change the shared config at the root. Package `tsconfig.json` files only `extends` the base and set `include`; `apps/*` add only what their framework needs (ADRs 0002, 0004). - Add a tool, a config file or an `apps/*` entry when real code needs it, and record why in a new ADR. - Secrets live in `.env` files, which are gitignored. Commit a `.env.example` with names only.

Skills

Three pairs of eyes

Real state, not guesses

A nightly agent reads production

Changes gated by a benchmark

Pre-registered, then measured

Want this way of working on your team?

Skills

Three pairs of eyes

Real state, not guesses

A nightly agent reads production

Changes gated by a benchmark

Pre-registered, then measured

Want this way of working on your team?