I'm rebuilding my whole site with AI, in public (part 1)

I decided to rebuild my website with Claude Code and document everything: the plan, the checkpoints, the wrong turns, the numbers. This is part 1, the setup and the first phases. The site you're reading is the result so far.

What you'll have when you're done

• The three-document setup that makes an AI rebuild survivable, ready to copy
• The real phase map from this rebuild, with what each stop caught
• The two moments where the plan met reality and reality won
• The sanitized version of my actual overhaul plan as the download
• [SCREENSHOT: the PROGRESS.md checkpoint log from this rebuild]

Before you start

This is a field-notes guide: follow along, steal what's useful. If you want the copy-paste method itself (the phase plan template, the tracker, the checkpoint prompts), that's the phase method guide. This series is what the method looks like when it collides with a real project.

Step 1: The three documents that ran everything

Before any code, I wrote three documents. Every session of the rebuild started by reading them.

• The master plan. One file with the whole build split into six phases, each ending in a hard STOP for my review. The prompt at the top says it plainly: "We build in phases. One phase per session. Never continue into the next phase without approval."
• The progress tracker. PROGRESS.md, the AI's memory between sessions: phase statuses, a checkpoint log, and a conflicts section with a standing rule I'd tattoo on every AI project: if a doc conflicts with the code, flag it, don't silently pick one.
• The brand spec. Every visual decision (colors, type, layout energy, image style) in one file, so design choices get made once and referenced forever, instead of re-argued with the AI every session.

Step 2: Phase 0, the audit that rewrote the plan

Phase 0's only job was to read everything and write down reality. No code changes allowed.

It mattered immediately. The audit discovered the project was further along than the plan assumed: a whole admin app I'd been building separately was already merged into the site in an earlier work stream, which meant two of the plan's phases were already done and two others were half-done. The plan got re-sequenced on day one, before any duplicate work happened.

That's the audit's entire value: ten minutes of reading versus weeks of rebuilding things that already existed.

Step 3: Phase 2, the taste checkpoint

The phase I was most nervous about: making an AI-built site not look AI-built.

The move that worked: instead of asking for "a nice design", the plan forced three complete visual directions, built as real comps on a hidden preview page. A warm cream print-zine direction, a warm charcoal direction with ember-orange accents, and a wildcard. Same content in all three, so only the design varied.

Then I picked, and here's the part worth stealing: I picked a mix (one direction's canvas and mood, the other's print hardware), and the winning tokens were locked into a config file, brand.config.json, that every later phase had to read. Colors, radii, shadows, type stack: decided once, in one file, no drift.

Step 4: The stops that saved it

The method's real product is the checkpoint log. A few entries from mine, verbatim in spirit:

• A review round where two brand logos turned out to have opaque backgrounds baked in; caught at the checkpoint, fixed with an image script, not shipped.
• A generated-identity experiment that got rejected at review because the renders drifted from how I actually look; the approach was replaced with reference-image generation using my real photos.
• A hero illustration cut entirely at click-through review because it read as filler.

None of those are code problems. They're taste and judgment problems, which is exactly what the STOP points exist to catch. The AI ships speed; the checkpoints ship standards.

Step 5: Where the series goes next

As of part 1, the rebuilt site is live: new positioning, a guides system with a publishing pipeline, gated downloads, and the brand applied everywhere. The upcoming parts cover the machinery built on top: the asset pipeline that produced the cover images (including the one on this page), the content systems, and the launch numbers, real ones, once there are enough weeks of data to be honest.