The platform is live. A private workspace for every engagement, and an agent that answers from your project or says it does not know.

Built before written about

The Lab.

Essays that end in a prompt you can run on your own repository. Agent architectures with demos you can play with. And the tools I run my own engagements on.

For your repository

Four prompts, in the order I would run them.

Each one is an essay that ends in a prompt. Paste the prompt into Claude Code, Cursor or Codex and it does the work on your repository: lays it out, brings it up to the standard, installs the rules an AI codes under, then audits the result.

Folder layout4 shapes

Pick the shape before the code

Four folder layouts, every folder with a role and a rule for what it may import. The prompt lays your repository out this way.

Read it, then run it

Production standard7 parts

Bring this repository up to the standard

Seven things a production codebase has to have, with the failure behind each. The prompt audits yours and installs what is missing.

Read it, then run it

Rules for AI coding5 files

The five files that run this codebase

The five files that let an AI write code here every day without the rules drifting. The prompt installs them in your repository.

Read it, then run it

Trust audit method40 checks

What I actually check

The checks I run on a codebase before I trust it, with the anti-pattern behind each. The prompt runs them on yours and reports back.

Read it, then run it

Agent architecture

How I build with agents, with demos you can play with.

Why naive RAG fails in production and the pipeline I run instead, why one agent hits a ceiling and the graph I draw instead, what a production agent harness is made of, and the decisions I make on real systems, as a drill.

RAG architecture6 demos

An agent proposes. A verifier and a human decide.

Why naive RAG fails in production and the pipeline I use instead: retrieve, verify, gate. Six demos from a real build.

Open it

Multi-agent architecture6 demos

Graph engineering: draw the shape of the work

One agent runs in a straight line. This runs a graph: fan out, verify, converge, and a loop that stops. Six demos.

Open it

Harness architecture6 demos

Harness engineering: what a good agent harness is made of

The parts of a production agent harness drawn as one architecture, with the trade-offs of each, and a prompt your agent runs to audit and improve yours.

Open it

Drill24 scenarios

Agentic AI in production

Real architecture scenarios. Guess the answer, then see mine with a diagram. Reveal, drill, shuffle.

Open it

Tools

What I run my own engagements on, opened up.

Built for my own work first. Both answer from cited facts, or say they do not know.

Product6 surfaces

The Platform

A private workspace for every engagement: documents, repositories, decisions, and an agent that answers from them with citations or refuses.

Open it

MCP server9 tools

Talk to my site

Add it to your AI client and ask what I do. It answers from cited facts and refuses when it does not know.

Open it

Not open yet: Trust Audit. Describe your AI system, or point at a repository, and get an honest read on whether it can be trusted in production.

Building something where this matters?

This is how I think on real systems. If you want that judgment pointed at yours, let us talk.

Let's talk