How I build

How I build with AI.

I write the spec, AI coding agents write most of the code, and nothing merges until automated checks pass and I’ve reviewed it. That loop is how one person ships a whole product, and it’s the same loop I run on client work.

  • Claude Code
  • Codex
  • Cursor
  • Figma
  • GitHub Actions

The loop

  1. Design and spec

    User testing, Figma designs, and motion demos run in parallel. The output is a written spec: what the feature does, the edge cases, and how we’ll know it works.

  2. Build with agents

    I hand the finished spec to Claude Code, my main tool. Each change lands as its own pull request. Every repository carries an AGENTS.md and shared agent rules, so agents follow the same conventions every time.

  3. Check

    Lint, type checks, contract checks, and tests run on every pull request. If any of them fail, it doesn’t merge, no matter who or what wrote the code.

  4. Review and ship

    I read every diff before it merges. Then it deploys. On Autodemia this loop ran more than 260 times in the first week.

How I check what AI writes

Speed only matters if the code holds up. These are the gates.

  • Tests gate every merge

    Men of Hunger has more than 11,000 automated tests across its API, web, iOS, and Android apps. Autodemia has more than 1,000, plus end-to-end browser tests.

  • A mock AI mode for tests

    Autodemia can run with the model swapped for a mock, so tests cover the AI features without calling a model or spending money.

  • Human review on every diff

    Agents propose; I approve. Work an agent starts on its own in Men of Hunger waits for my approval before it goes out.

AI inside the product

I also ship AI features and agents that customers use.

  • Assistants that call tools

    Marv, the assistant in Men of Hunger, answers in threads, DMs, and group chats. It calls 18 tools, picks from three model sizes, and keeps private conversations out of public answers.

  • MCP servers

    Men of Hunger has an MCP server and command-line tool with about 70 tools, so AI agents can read analytics and draft posts that wait for approval. Autodemia has its own MCP endpoint too.

  • Scoring answers every night

    A nightly job scores Marv’s answers and publishes lessons as a versioned playbook I can roll back. It improves without retraining.

Keeping AI costs in check

AI features have a running cost. I design for it from the start.

  • The cheapest model that fits

    Autodemia starts each call on the smallest Claude model that can do the job and steps up only when you correct it.

  • Credits and caps

    Credit caps keep Autodemia’s AI spend below revenue. Men of Hunger gives members an allowance.

  • Cost tracked per call

    Every AI call’s cost is recorded, and an alert fires when a day’s spend triples.