Process
How I build with AI
I don’t write most of the code on my builds. I decide what gets built, direct the models that build it, catch what they get wrong, and ship it. The model matters less than the process around it.
The workflow, in order
-
Scope with the end user.
Write the problem in their words before touching a tool.
Example: the leasing system was scoped with the team that uses it every day.
-
Plan in one model, build in another, review in a third.
Separate models catch each other’s blind spots.
Example: on TRS, ChatGPT planned, Codex and Claude Code built, and Claude reviewed.
-
Govern the build.
Checkpoints, decision records, CLAUDE.md and AGENTS.md files that keep each project’s rules explicit, and reviewed PRs.
Example: the leasing system’s decision records, D1–D17, document every trade-off.
-
Verify before trusting.
Every field a human acts on needs a source and a validation step.
Example: the job search agent’s validation gates.
-
Ship small, measure, expand.
Expansion only after the first thing is used.
Example: TRS started with one reconciliation tool.
Two judgment calls
Not automating too early
In mid-2026 I evaluated n8n for orchestrating Polaris, my governed workflow for AI-assisted builds, and chose to wait until the manual workflow was stable. Automating an unstable process locks in its problems.
Auditing my own production code before publishing it
Before making a public demo of the leasing system, I ran a full read-only audit of it. It found five pricing disagreements between the two apps, five behavior bugs, and open database rules. The fixes are being made in the demo, test first, and the same fixes go to production next. Publishing the cleaned version with the audit summary is the point.
Tools
- Claude
- Claude Code
- Claude Cowork
- Anthropic API
- ChatGPT
- OpenAI Codex
- GitHub
- Firebase
- Supabase
- Netlify
- Cloudflare Pages
- HTML/CSS/JavaScript
- Python
- Excel and Power Query
Only tools I’ve shipped with.