Skip to content

/work · LLC · Portland, ME gym

Built · launch anticipated September 2026

CrossFit MF / Misfit Gym Portland

A Portland, Maine gym (est. 2009) gets a full marketing-site rebuild, and the build itself is the best-documented example of my agent workflow end to end.

01 · Context

CrossFit MF (Misfit Gym) has operated in Portland, Maine since 2009. A local gym's website has one job: convince a nervous first-timer to walk in the door. That makes it a marketing problem with ethical edges, which is why the docs package carries explicit honest-CTA rules alongside the design direction.

02 · What I built

A complete marketing-site rebuild:

  • Purple/teal token-based theming with full light and dark modes
  • Signature components: a 10-physical-skills accordion gallery, a weekly schedule grid, and member stories
  • Honest-CTA rules applied throughout, so the site persuades without manufactured urgency
  • Local SEO done structurally: LocalBusiness schema and locally-scoped metadata
The hero: real members, real rating, honest CTAs.01 / 03

03 · The senior-engineer part

  • Specifying the build completely enough for autonomous execution: 14 documents covering brand, voice, verbatim content, requirements with decision-authority rules, design direction, IA, components, SEO, and CTA ethics
  • A token-based theme architecture so purple/teal, light/dark, and every component stay consistent by construction rather than by vigilance
  • Treating CTA ethics as a requirement with rules, not a vibe: deciding in writing what the site will and won't do to get a signup
  • Local SEO done structurally: LocalBusiness schema, not just keywords

04 · How AI was used

This is the star section, because this build is the fully documented one. The site was produced by the documentation-first workflow: I wrote the 14-doc package and the autonomous kickoff prompt; Claude Code decomposed the build into subagent workstreams, executed against the docs, looped on quality gates (build, typecheck, lint, accessibility, performance, both themes), and maintained OPEN-ITEMS and PLACEHOLDERS files for anything it couldn't verify.

The division of labor was explicit: the docs' decision-authority rules said what the agent could decide and what came back to me. I made the judgment calls (brand, ethics, content) before the build started, and reviewed the output against the package after. The full method, with this project as Exhibit A, is walked through on /systems.

05 · Stack