The making of

This wasn't vibed out in an afternoon.

Nine days. 241 commits. Two AI orchestrators, a block theme, fifty-five patterns, and twelve real businesses built to prove it. Here's the honest archaeology — the pivot, the rollback, the spike we scored against ourselves, and the testing that earned the result.

241commits in 9 days
15tagged releases
55block patterns
12real businesses built

The short version

It started as another WordPress theme. By day three it had become something nobody had shipped: the compiler that turns a one-paragraph brief into a real, editable WordPress site.

Most "AI makes you a website" tools are a thin wrapper over one model call. AgentPress is the opposite: a planner that decides the sitemap, writers that produce brand-voice copy, an image engine that keeps a consistent illustrated style across a page, a palette compiler that guarantees legibility by construction, and a pattern library that renders it all as native, hand-editable WordPress blocks. The interesting part isn't any one of those — it's the nine days of choosing, testing, and rebuilding that made them work together.

The build log

Nine days, day by day

Pulled straight from the commit history and the dated working notes. Nothing dramatised.

Day 1 — 26 May

It started as "another block theme"

v1.0.0 to v2.0.0 in a single day. Three style variations, a contrast-safe palette, per-variation headers and footers. The discipline showed early: v2.0.0's headline was a "design-system overhaul," not "add more features."

Day 2 — 27 May

The first hard lesson, and the first rollback

We built a pattern library and a scaffolder, then hit a wall: WordPress's full-width layout kept fighting our hand-authored markup. The note in the task list reads, word for word: "STOP and assess."

So we did. We read Automattic's own block-theming guidance, archived two versions of working-ish code, and rolled back to rebuild the whole pattern library on their verified foundation. Throwing away progress on day two to stand on firmer ground is the opposite of vibe-coding.

Day 3 — 28 May

The pivot that changed the product

We stopped and asked a bigger question. The answer: the market doesn't need another opinionated theme — it needs the WordPress destination for the emerging DESIGN.md ecosystem. We rewrote the plan around a compiler, and committed to a rule we still hold: invent no new vocabulary. Material Design's token names pass through unchanged, end to end.

Day 4 — 30 May

The road not taken, scored honestly

The busiest day — 52 commits. We spiked an alternative "match a rendered design to patterns" pipeline and set the bar before running it: clear 80% agreement with a hand-graded standard, or don't build the full corpus. It hit 83%. And when an early draft of the results overclaimed, the correction was written into the permanent record rather than quietly edited away.

Days 5–6 — 31 May / 1 Jun

Resilience and real content

Retry-and-fail-loud builds, multi-level site structure, a contrast validator that guarantees readable text by construction, and grounded content — sub-pages, suburb pages and blog posts written from real search data, never invented facts.

Day 7 — 2 Jun

Building the test bench

Stood up a dedicated iteration box and a dozen WordPress installs so we could rebuild whole sites in minutes and watch what actually changed — capturing each gotcha as a rule so it never cost us twice.

Days 8–9 — 3 Jun

The "designed-for-us" craft

Consistent illustrated process sets instead of unrelated stock, the client's real logo composited onto a uniform in a signature shot, editorial-rhythm sections, sticky calls to action — then a browse-every-page audit of all twelve demos, fixing root causes, not symptoms.

Proof it was deliberate

The roads we chose not to take

Every one of these was tried, measured, and set aside for a reason. That's where the judgement lives.

Pivoted

Another opinionated theme

The obvious first product. Set aside on day three — the market needs a compiler, not one more theme.

Rebuilt

Hand-authored markup

Full control, but it fought WordPress. Replaced with Automattic's verified pattern shape so it never drifts.

Gated

HTML-first design matcher

Could ingest any rendered design. Spiked to 83%, kept as a validated option — the simpler blueprint-driven path won.

Refused

Our own token names

Tempting branding. Deliberately rejected — Material Design's vocabulary stays intact end to end.

Banned

Placeholder phone numbers

A fake "0400 000 000" on a real business is worse than none. Empty-unless-real is enforced at the source.

Instrumented

Eyeballing "looks fine"

Replaced with scripts: contrast computed at the source, duplicate-image guarantees, live end-to-end checks.

What we stood on

We invented almost nothing

The craft was in the integration, not in novel primitives. AgentPress consumes published standards and plugs into the ecosystem around it.

  • DESIGN.md — the brand-brief format, read as published
  • Material Design 3 — the colour token vocabulary, unchanged
  • W3C DTCG — the token export format
  • WordPress MCP + Abilities — how agents drive the site
  • Automattic block-theming — the drift-free pattern shape
  • Material Color Utilities — legible palettes by construction

How we knew it was good

Testing, not vibes

Quality was measured, not assumed.

  • Brains-trust panels — three frontier models reviewed live screenshots; the coffee demo went through nine rounds
  • Gates before corpora — the design spike had to clear 80% before earning a bigger build
  • Instrumented dogfooding — scripts caught what eyes miss
  • Live verification — "done" meant the real URL worked end to end
  • Twelve real businesses — rebuilt repeatedly as the engine improved

So — try it, and judge the result yourself.

A one-paragraph brief becomes a complete, editable WordPress site. The nine days are why it doesn't look like everyone else's.