1. Start with the need, not the tech
My trigger wasn't "I want to build a site." It was: "I'm tired of re-explaining what I do with AI — I need one single place." The need dictated the form.
2. Choose the dumbest possible stack
- Site: vanilla HTML/CSS/JS, zero framework, zero build. Open an
.html, it works. GitHub Pages hosting (free, deploys on push). - Bot: a Python script orchestrated by GitHub Actions (a free cron). No server to manage.
3. Separate responsibilities
Two repos: bot = data engine, site = display. Connected by one single file (feed.json). Immediate benefit: secrets (API keys, bot token) stay in the private engine repo — the public site contains none.
4. Work in a visual loop with AI
My rhythm: I describe → AI codes and commits → I look at the render in the browser → I ask for an adjustment. In a loop. AI holds the keyboard; I hold the choices (tone, structure, safeguards) and the verification (accurate facts, nothing invented).
5. Set safeguards from day one
- Budget: 1 run/day, strict caps on collection volume, no runaway loops.
- Copyright: never copy-paste content — always a reformulated summary + the source link.
- Security: never publish a secret (placeholders everywhere), one dedicated token per use.
Key takeaway. The bottleneck is no longer "I don't know how to code." The work that remains — deciding, scoping, verifying — is exactly what a business owner or retailer already knows how to do.