What happens if you give Claude Code an ugly website and say nothing but "make it better", over and over? I built the ugliest coffee shop page I could (a scrolling banner, Comic Sans, a broken photo, typos, 28 lines of HTML), then ran the same prompt 10 times on Sonnet 5.5 and 10 times on Opus 5.5. Every round was a fresh session with no memory of the round before, and I saved a screenshot and a full copy of the site after each one. These are our own numbers from that run on 2 November 2026. The presenter in the reel is an AI avatar of Hammad; the words and the experiment are his.
The prompt, sent every round
Each round got the same five instructions: open the website first and look at it, then innovate; put yourself in the customer's shoes and think beyond the obvious; make one bold change that lifts conversions or the look and feel, not a minor tweak; use any tool you have; and you must ship a change this round. The last line matters. Without it, a loop like this can decide the site is finished and stop changing anything, so forcing a shipped change every round is what keeps the experiment moving.
Sonnet 5.5: careful, ordering-first upgrades
Sonnet's first round replaced the yellow page with a warm, modern redesign and an open-now badge built from the real opening hours. After that it stayed close to one goal: getting people to order. It added an order-ahead flow that texts the order to the shop, a one-tap "order my usual", a pastry upsell, a time-of-day pick, a two-tap taste quiz and illustrated menu cards. In round 3 one change broke the stamp card and the hours list, and round 4 noticed and fixed it. It finished at 231 lines after about 12 minutes of agent time across the 10 rounds.
Opus 5.5: it kept inventing, then made things up
Opus wrote more from the start (272 lines after round 1) and never slowed down: a build-your-drink sheet with a live illustrated cup, a three-tap coffee matchmaker, illustrated menu cards with badges, a group coffee-run order, whole-bean bags to take home and a gift-a-coffee flow with a shareable link. It finished at 1,089 lines, almost five times Sonnet's, after about 23 minutes of agent time. The moment it went too far was round 9: it added an "order like a regular" section with three customer reviews, named regulars and a "4.9 average from our regulars" rating, none of which existed. On a real business site, invented reviews are a liability, not a feature.
Zero web searches in 20 rounds
ONE TACTIC A WEEK
The prompt told both models to use any tool they had, and web search was available to them, yet neither searched the web once in 20 rounds. Every idea came from what the model already knew. That matches what I see in my own automations: the best output comes when I hand the AI a proven example to work from, not a vague goal. If you want inspiration from award-winning sites or a component library, you have to point the model at it.
Opus 5.5 vs Sonnet 5.5: which should you use?
For a loop like this, Sonnet 5.5 was the safer pick: smaller, focused changes that stayed on one goal, at half the per-token price (Sonnet 5.5 is $2 in and $10 out per million tokens on the API, Opus 5.5 is $4 and $20, per Anthropic's pricing page on 2 November 2026). Opus 5.5 is the one to use when you want bold new ideas, but give it limits, review every round, and check that nothing it adds claims facts about your business that are not true. Either way, my guess is that "make it better" gets you to about 80 out of 100. The last 20 is taste, and that still comes from you.
Run it yourself, safely
Run the loop on a copy of the site, never on a live one, and commit after every round so any change can be rolled back. In headless mode (claude -p) you can cap a run with --max-turns and --max-budget-usd, but both are estimates and can overshoot. Claude Code is not free and not open source: it needs a paid Claude plan (Pro, Max, Ultra, Team or Enterprise) or an API key.
Get the exact prompt
The free guide at withhammad.com/resources/make-it-better-loop has the exact prompt to copy, how the run was set up, the results and the safety rules. On Instagram, comment BETTER on the reel and the guide link arrives by DM (follow @withhammad.ai first so the DM reaches you).
Related reading
If this is useful, see how to reverse-engineer a feature with Claude Code and the brief that stops AI websites looking generic.