I didn’t write most of the code by hand. I directed the work using Claude Code, Anthropic’s CLI tool for AI-assisted development. Here’s what that process actually looked like, because the reality is more nuanced than “the AI wrote it for me.”
Saturday morning: research and fact-checking
I’d seen the Kindle story in a Claude browser session earlier in the week, but browser Claude doesn’t have tools. It can’t verify its own claims. My first job was checking whether the “May 20” date was even real, or whether the AI had hallucinated it.
Claude Code launched three parallel research agents: one to verify the cutoff facts against Amazon’s help page and press coverage, one to validate every source URL cited in the earlier research, and one to check domain availability.
Results came back in under 15 minutes. The date was real, confirmed by TechCrunch, HowToGeek, TechRadar, T3, and Engadget, all published 8 April 2026. Twenty-five of twenty-seven source URLs were genuine. One had fabricated content. Two were real but returned 403s from bot-blocking.
That research sprint gave me confidence to proceed. Without it, I’d have been building on sand.
Saturday afternoon: site architecture and first pages
I described what I wanted: a consumer-help site with free content, a paid PDF guide, and Stripe checkout. Claude Code built the directory structure, shared header/footer includes, a custom CSS theme, and the first three content pages in a single pass.
The important thing: I reviewed every page before moving on. The AI wrote the first draft; I read it as a consumer would. Where the tone was wrong (too technical, too salesy, too vague) I said so and it adjusted. This is the part that takes taste, not tools.
By Saturday evening I had 13 free pages including a “Do Not Reset” warning (the single most important piece of consumer advice for affected Kindle owners), a full explainer, a three-options overview, and four detailed repurposing guides.
Sunday morning: the paid product
The guide started as a 17-page PDF with one-sentence summaries for each of the 20 repurposing projects. When I asked “tell me about the Minecraft server instructions”, there weren’t any. Just a sentence and a link. That wasn’t good enough for a paid product.
So we expanded Chapter 5 from 800 words to 13,000. Every project got a proper how-to: what it does, what you need, numbered steps, gotchas, and source links. The content went into a shared PHP include so the online edition and the PDF stay in sync automatically.
Two independent editorial reviews then caught a critical safety omission (disabling firmware auto-updates before jailbreaking) and suggested a dozen improvements including a quick-start page, a glossary, and a printable “DO NOT RESET” sticker sheet. All were implemented in one pass.
The PDF is rendered from HTML via headless Chromium, compressed with Ghostscript. 56 pages, 1.5MB, with 28 embedded illustrations.
Sunday afternoon: images, checkout, deployment
I wrote DALL-E prompts for 33 illustrations, all following a consistent style brief with the site’s warm ochre-and-navy palette. Generated them in ChatGPT, downloaded, and Claude Code processed them to WebP, resized, and placed them in the correct directories. Every image was reviewed before acceptance; one was regenerated because of text bleed from a prior prompt.
The Stripe integration uses Stripe Checkout with automatic UK VAT via Stripe Tax. Access tokens are HMAC-signed URLs with no user accounts, no passwords, and no database. A refunded order immediately revokes access. The whole payment system is about 200 lines of PHP.
Deployment is rsync over SSH. The site runs on Apache with PHP 8.3. The config auto-detects whether it’s running on the dev host or production and selects the correct Stripe keys accordingly.
Sunday evening: testing and SEO
Claude Code set up Playwright in the project and wrote 58 end-to-end tests covering every page load, every navigation link, image loading, Stripe form presence, access control (token-gated pages return 403 without valid tokens), SEO files, price consistency, responsive layout, and domain consistency checks. All 58 passed.
Eight SEO landing pages target model-specific searches (“Kindle Keyboard after May 2026”, “Kindle stuck on registration screen”) with FAQ schema for Google rich snippets. The site was submitted to Google Search Console and IndexNow on go-live.
What I actually did vs. what the AI did
This is the part people always want to know.
The AI wrote the code, the HTML, the CSS, the content first-drafts, the Stripe integration, the access-token system, the test suite, the SEO files, the image processing pipeline, and the PDF generation toolchain.
I made the decisions: what to build, who it’s for, what the tone should be, what price to charge, which facts to trust, which images to accept, which content was good enough to sell, and when to push back. I also handled the Stripe dashboard, domain registration, DNS, and deployment configuration.
The ratio is roughly 90% AI-generated output, 100% human-directed. Nothing shipped without my review. The AI never decided what to build. It built what I asked for, and I asked for it in detail.
The result
oldkindle.com is live. Free consumer help for anyone with a pre-2012 Kindle. The “Do Not Reset” warning page alone is worth sharing. People are genuinely bricking their devices on bad advice from Amazon’s own chat support.
The paid guide is £3.99: 56 pages covering the pre-cutoff checklist, Calibre sideloading, WinterBreak jailbreaking, KOReader, 20 repurposing projects, troubleshooting, and a model compatibility matrix. It won’t make anyone rich, but the marginal cost of serving it is zero and it saves affected owners a genuine hour of searching.
This was built over a weekend by one person with an AI coding tool and a domain name. The interesting question isn’t whether AI can build a website; it obviously can. The interesting question is whether a solo practitioner, working with AI, can research a breaking story, fact-check it, build a product around it, and ship it to production before the news cycle moves on.
The answer, this weekend at least, was yes.
The numbers
- 23 pages of free content
- 8 SEO landing pages with FAQ schema
- 56-page paid PDF guide with 28 illustrations
- Stripe Checkout with automatic UK VAT
- HMAC-signed access tokens (no database)
- 58 Playwright end-to-end tests, all passing
- Google Search Console + IndexNow submitted
- Total development time: two days
- Total cost: domain registration + Stripe fees on sales