PressBot.io got a full redesign last weekend. Twenty-two pages rewritten from scratch, a new type system, site-wide light and dark mode support, with light mode being the new default and a production cutover that fit into a single afternoon. The tool was Claude Opus 4.7 running in the Claude Desktop app.
The interesting part of this story isn’t “three days.” Everyone has a three-days story now. The interesting part is what the model demands from you to get there — and why that turns out to be good news, not bad.
What’s New on PressBot.io
If you haven’t been by the site in the last week, every page you land on is v2. A few worth poking at:
- The new homepage — less hero theater, more honest pitch.
- The tour — real screenshots in a lightbox, not polished product mockups.
- Tidio alternative and Crisp alternative — each with a “Recommended for / Skip if” split that tells you when PressBot isn’t the right answer.
- Pro — BYOK pricing math spelled out in plain numbers instead of buried in an asterisk.
- Light and dark mode site-wide, toggled in the header. No flash, no repaint, your preference sticks.
Behind the scenes: 22 canonical pages plus four filename-driven templates (front-page, single, 404, index). One rsync to push the v2 templates to production, then 22 WP-CLI commands to flip each page from v1 to v2. Non-destructive — v1 is still on disk, so every page can be rolled back in one click. That’s not glamour. It’s discipline. Discipline is why nothing broke.
Anthropic Told Developers to Re-Tune Their Prompts
When Anthropic shipped Opus 4.7 in April, the release notes carried a warning most upgrade posts glossed over:
“Opus 4.7 interprets prompts more literally and explicitly… It will not silently generalize an instruction from one item to another, and it will not infer requests you didn’t make.”
Translation: prompts that were vague but “good enough” on older models now produce worse output, not equivalent output. The model takes your hand-waving literally, delivers exactly what you asked for, and you figure out after the fact that what you asked for wasn’t what you wanted.
Which means you can’t coast.
You can’t drop a half-formed idea into the chat and get the other 60% for free. If you want Opus 4.7 to build you something specific, you have to say what “specific” means.
What That Looks Like in Practice
Here is the contrast that mattered most across the redesign.
Prompt that used to work, doesn’t anymore: “Make the pricing section nicer.”
Prompt that works now: “Rewrite the pricing section on the Pro page. Keep the three-tier layout. Replace the feature grid with a ‘Recommended for / Skip if’ split, same visual weight as the current comparison table. Use the existing body-text token for the skip-if copy. Preserve the JSON-LD block at the bottom.”
Same goal. Very different outputs. The first prompt gets you a generic prettier gradient. The second one gets you the actual design choice, on brand, in one pass.
Multiply that across 22 pages, and you see why the redesign was possible in a weekend instead of a month — and why it required more thinking from me, not less.
Why This Is Actually Good News
The anxious version of the AI-coding story is that the model will get so good that humans become redundant. Opus 4.7 suggests the opposite.
This is not a model that makes your thinking optional. It’s a model that rewards your thinking — with output that’s roughly ten times what you’d produce alone, provided you brought real thinking to the exchange. If you don’t bring it, the output quality collapses, and it collapses faster than it used to.
It feels less like delegating to an AI and more like pair programming with a sharp colleague who won’t accept a lazy pseudocode sketch as a spec.
The model that punishes laziness and rewards care.
That’s the version of AI-assisted work we actually want. Not a magic box that replaces you. A capable counterpart that’s only as capable as you make it.
Four Rules That Made the Redesign Possible
If you’re considering an AI-assisted redesign — of your own site, a client’s site, a plugin landing page — here’s the tactical shortlist that held up across 22 pages:
- Write specs, not wishes. Every prompt should name: which file, which pattern to preserve, which tokens to reuse, what the output should look like, and what not to change. If you can’t name those, you’re not ready to prompt yet.
- Pick the tool that matches the shape of the work. For a redesign — where “what I want” is aesthetic and editorial — a chat interface like Claude Desktop beats an autonomous terminal agent. For refactors and test suites, the terminal probably wins. The tool is not the point. The cadence is.
- Deploy non-destructively. Keep v1 and v2 side by side on production. Flip at the template level, not the file level. Rollback should be a single click, not an incident.
- Walk the rendered page. No model catches every visual edge case from the inside. Load the site in a real browser, in both themes, and write down every spot that looks off. Then re-prompt with those notes. That second pass is where the quality lives.
See the Result
Take a look around. Toggle the sun/moon icon in the header. Open the Tidio comparison page in dark mode. Check the real screenshots on the tour. Read the “Skip if” block on Pro and notice how much more trustworthy a page feels when it tells you when not to buy.
Then, the next time you sit down with Opus 4.7, bring the spec before you bring the wish. The model will meet you where you are — and lift the output from there.
