how-i-build
Agents in the loop, every working day.
Everything on this site was shipped with AI in the loop — ShimmerWorks, Breakfast Ball, the site itself. Not autocomplete: agent harnesses that scope, build, ship, and audit, with me steering. This page is the method.
- Tokens · 2025
- 2.59B
- Agent runs
- 6.3K
- Custom skills
- 17
- Measured, not assumed
- A/B
through the loop
2025 alone
authored · ~/.claude/skills/
evals · error bars reported
the-loop
Five steps. The harness compounds.
The same loop shipped a sold SaaS platform and a watch app in beta. It isn't about typing less — it's about closing the distance between deciding and verifying.
01 / Scope
The operator brain frames the actual job. No agent starts until the goal fits in one sentence.
02 / Spike
Claude Code cuts the first working version against a fast loop. Failures are visible within minutes.
03 / Harness
Custom skills and structured prompts turn one-off spikes into repeatable patterns. The harness compounds.
04 / Ship
Type-check, build, push. CI deploys. The loop closes end-to-end, every time.
05 / Audit
Adversarial review passes catch the class of mistakes humans-with-agents make.
the-library
Skills I've shipped. Every one is a tool I actually use.
This site was built by site-rebuilder, deployed by ship-site, and audited by refine-rebuild — all custom Claude Code skills. A sample of the library:
site-rebuilder
End-to-end pipeline for scraping a live site and rebuilding it as a mobile-first SvelteKit + Tailwind site. Built this site.
ship-site
Takes a built SvelteKit project, creates a private GitHub repo, wires Cloudflare Pages + GitHub Actions, verifies the first deploy is green.
refine-rebuild
Closing-pass audit for the mistakes rebuilt sites quietly ship with — misidentified photos, mobile overflow, missing OG metadata, fabricated facts.
resume-builder
Ingests a folder of messy career material and produces a master resume plus role-tailored persona variants in PDF + HTML.
llm-council
Runs any decision through a council of 5 AI advisors who independently analyze, anonymously peer-review, and synthesize a verdict.
placerville-coop-minutes
Generates board meeting minutes from audio transcripts. Standardized, legally-disciplined PDF + HTML. Real civic tooling, in production.
hosting-investigator
Deep-dive on who hosts a website, who registered the domain, and what the owner is paying — including renewal-trap math.
website-auditor
Builds a website audit dashboard for small businesses — dead links, accessibility, SEO, UX, security flaws.
living-design-system
Turns an existing codebase into an interactive component-level design system with adjustable playgrounds and an annotation layer.
harden-secrets
Moves API keys out of plaintext into macOS Keychain and hardens the agent harness against credential leakage — denylists + redaction hooks.
work with me
If you want this loop pointed at your problem —
Strategy Labs takes on a small number of client builds. Email is the fast path.
stack
The tools, by era.
The operator-era tools earned their place — a decade with them — but the new stack is what I reach for first.
01
AI & agents
Daily practice, not a bullet point. 17 authored agent skills, measured with A/B evals rather than assumed.
- Claude Code
- Claude API
- Codex (OpenAI)
- Cursor
- Custom Claude Code skills (17 authored)
- Custom agent harnesses
- RAG pipelines
- On-prem inference (CUDA · Apple Silicon)
- Voice (Whisper · TTS)
02
Web & data
What ShimmerWorks shipped on, and what Strategy Labs ships on now.
- TypeScript
- Svelte 5 · SvelteKit
- PostgreSQL (bitemporal · RLS)
- Supabase
- WebGL · GLSL
- Node.js
- Python · FastAPI
- Tailwind CSS
- WebSockets
- Vitest · Playwright
- AWS · Cloudflare
- GitHub Actions · Sentry
03
Native & mobile
Breakfast Ball's stack — four Apple targets, in TestFlight beta.
- Swift · SwiftUI
- ActivityKit (Live Activities)
- WidgetKit
- WatchConnectivity
- CoreLocation · MapKit
- Capacitor (plugin author)
- Kotlin services
- APNS
- TestFlight · App Store Connect