How I build with AI

AI accelerates. Judgment is not delegatable.

The loop is always the same: plan first — the model gets a spec, not a vibe. Scaffold fast — generation is where AI genuinely multiplies pace. Review like it's a stranger's PR — because functionally, it is. Writing code stopped being the bottleneck; deciding what code should exist, and whether this code is that code, is the job now.

Tools change. Standards don't. The verification bar — spec-checked accessibility, tested edge cases, honest metrics — is identical whether the first draft came from my hands or a model.

Receipts, not vibes

Examples from product work and independent prototypes. Each names what the AI did, what I did, and which artifact you can inspect.

• A reviewed script becomes a product walkthrough

  Waco3.io · video workflow in active development

  I review the scripts, then use Strands Agents to coordinate screen capture, Kokoro.js narration, composition with HyperFrames or Remotion, and validation. The post explains the responsibilities and includes a TypeScript scaffold. Read the workflow →

• Primer guidance, checked against the implementation

  Agent Review Desk · independent prototype

  The local Primer MCP server supplied component APIs, token guidance, and accessibility notes. A broad token search returned no result; narrowing by semantic group made it useful. Installed types, the state model, and browser checks remained the verification steps. Case study →

• Bundle claims that can be reproduced

  ReactZero · four published packages

  The old portfolio used one size claim for several different imports. The replacement fixture pins released packages, externalizes React, and separates selected JavaScript exports, all public exports, and stylesheet cost. The resulting measurements are published with the script. Case study →

• An explicit review boundary

  Agent Review Desk · state model

  The prototype separates permission, check completion, and human review. Tests verify that denial cannot be bypassed and that failure expires approval. Generated scaffolding is useful only after these product rules are made explicit. Case study →

• Product behavior that preserves user work

  Waco3.io · working product

  The proposal editor keeps generated sections editable and treats partial output and retry as product states. I continue building the working product; usage exists, but this portfolio does not invent customer totals or business results. Case study →

Where the line sits

I also build for AI

Using AI well and designing for it are the same literacy. Every ReactZero library ships an ai-reference.md — structured documentation written for AI coding assistants, so when someone integrates a component through Claude Code, Copilot, or Cursor, the model generates correct, accessible code instead of a plausible guess. And Waco3.io is an AI product whose interface work — streaming as a design material, edit-in-place trust patterns, human-in-the-loop by default — is documented in its case study.

I think about this in public

A selection from the writing section — the workflow above, argued one piece at a time.

• How I use Strands Agents to make Waco3 tutorialsReviewed scripts, specialist workers, and a TypeScript scaffold.
• The best AI workflow ends with human reviewThe operating rule, argued in full.
• Agents should show their workWhy visible reasoning beats confident output.
• Hooks are where AI coding gets seriousGuardrails as configuration, not vigilance.
• How to give an LLM a design systemThe thinking behind ai-reference.md.
• Subagents are not employeesDelegation models that actually work.
• Claude Code skills: procedures with memoryEncoding workflow so it compounds.