AI Website Experiment

Status: Active. This is the running record of an experiment in building, correcting, and finishing a real WordPress website with AI collaboration.

Purpose

The question is not merely whether an AI can generate a web page. The useful question is whether a human and AI can carry a website through the entire construction process: discovery, design, content migration, navigation, responsive behavior, troubleshooting, verification, documentation, and eventual handoff.

Working questions

• Can an AI translate a broad aesthetic preference into concrete design decisions without constant micromanagement?
• Can the work remain coherent while the user steers, interrupts, refines, and changes levels of abstraction?
• Can the AI act decisively while still recognizing the few moments that genuinely require human confirmation?
• What does a genuinely completed AI-assisted website cost when AI usage, infrastructure, licenses, and human labor are measured separately?
• Can long conversations remain navigable without flattening them into summaries?

Design principle

Subtle yet spectacular: calm at rest, remarkable when attention finds the hidden detail, and never heavier than the technology or the reader can comfortably carry.

The visual system uses light ivory, restrained Flower-of-Life geometry, gold and silver Phoenix artwork, a compact reading-first layout, and small moments of motion that respect reduced-motion preferences.

Method

1. Build in a fresh, isolated staging installation.
2. Import the existing public library without cloning old structural problems.
3. Create a custom, free Phoenix theme around the content rather than forcing the content into a commercial template.
4. Provide persistent conversation navigation on long entries and a compact mobile drawer.
5. Test at desktop, tablet, and phone widths.
6. Record discoveries, corrections, costs, and failed routes as evidence rather than hiding them.

Experimental controls

• Separate observed facts, interpretations, and hypotheses.
• Use a hypothesis-signal log for genuinely new secondary-goal inferences. Report a timestamped addendum when three reasonably independent observations converge, or when one strong behavioral pivot supports a new hypothesis with a plausible competing explanation. Each addendum records evidence, confidence, alternatives, disconfirming evidence, and why the result is new.
• Distinguish “AI generated it” from “the system was completed and verified.”
• Prefer reversible staging changes and visible checkpoints.
• Evaluate completeness across content, design, accessibility, responsive behavior, performance, security, integrations, and handoff readiness.

Running log

August 21, 2026 — Foundation

• Created a clean WordPress staging installation at AnyKey Cafe.
• Corrected an early false start in which a clone had been created instead of a fresh installation.
• Imported the existing public library: 75 posts and 85 pages, including the model-specific categories.
• Built and activated the AnyKey Phoenix theme using only free and locally created components.
• Added the ivory, gold, and silver visual system; subtle geometric background; responsive Phoenix-first hero; reading progress; persistent conversation rail; mobile conversation drawer; and reduced-motion handling.
• Detected and corrected an oversized automatically imported navigation menu before handoff.
• Implemented the user-authorized Hypothesis Signal Log with four retrospective calibration seeds and a prospective timestamped reporting trigger.

The cost variable

Cost will be recorded in separate columns so the result cannot hide behind a single misleading total:

• AI operation: subscriptions, usage credits, tool calls, and image generation.
• Infrastructure: domain, hosting, storage, email, and backups.
• Software and assets: themes, plugins, licenses, fonts, photographs, and media.
• Human operation: decision-making, account recovery, support calls, content review, and final approval.
• Rework: false starts, failed integrations, external service faults, and changes of direction.

A cheap generated page and a completed website are different products. This experiment measures the second one.

This entry will be updated as the build progresses. Conclusions remain provisional until the final audit and handoff.

Additional variables added

Client Content Readiness Index

The experiment now records how prepared the human-supplied material is before construction begins. CR-0 means the material does not yet exist; CR-1 means it exists only as ideas or scattered notes; CR-2 means partial digital material exists but requires substantial assembly; CR-3 means most material is supplied digitally with moderate cleanup; and CR-4 means the material is organized, approved, and ready to place. This separates the cost of building a site from the often-hidden cost of discovering, collecting, rewriting, and approving its content.

Context-profile hypothesis

A candidate hypothesis is being tracked: giving an AI relevant working preferences, terminology, decision rules, and corrections before a complex task may reduce translation errors and revision cycles. This is not treated as a conclusion. It will be compared against task complexity, prompt quality, model differences, and the possibility that simple familiarity—not a formal profile—accounts for any improvement.

Observer-influence control

To reduce steering after the build begins, the observer can provide requirements first, avoid watching intermediate visuals, and review only at named reveal checkpoints. Changes made after a reveal are recorded separately. This does not eliminate observer influence, but it makes the timing and direction of that influence more visible.

Audit checkpoint — August 21, 2026

• Responsive behavior was checked at phone (375 × 812), tablet (768 × 1024), and desktop (1280 × 720) sizes.
• The long conversation hubs, the persistent conversation selector, the mobile drawer, search, category archives, tables, and the custom 404 page were exercised.
• The Phoenix artwork delivery payload was reduced from approximately 5.77 MB of source PNGs to approximately 299 KB of responsive WebP variants—a reduction of about 94.8%—while preserving the source artwork outside the deployed theme.
• Navigation destinations, the mobile administrator-toolbar overflow, and the search-results heading were corrected during the final pass.
• WordPress Site Health reported no critical issue. Its remaining recommendations were hosting-level performance options or routine cleanup items, not launch blockers for this staging experiment.

Public observation note

A link to this running record is now placed on the staging home page for readers arriving from the Farsight discussion. This page is a structured experimental log and checkpoint record; it is not represented as a verbatim transcript of every private message. A fuller transcript may be added later if it is deliberately prepared for publication.