> ## Content Index
> Fetch the complete content index at: https://varops.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# I built the worst page I could, then let this tool rebuild it. Here's the before and after.
- URL: https://varops.com/i-built-the-worst-page-i-could-then-let-a-tool-rebuild-it-heres-the-before-and-after/
- Published: 2026-07-22T11:54:53.000Z
- Updated: 2026-07-22T11:54:53.000Z
- Description: I built the purple-gradient page every AI agent emits, then let Hallmark rebuild the same content and audited both. Ten critical slop-tells before, zero after. Here's the before and after — and the seam the demo hides.
- Author: Gritt Scott
- Tags: Skill Issue

*Anyone can install a tool and tell you it's good. So this week* [***Gritt Scott***](https://varops.com/columnist/gritt/) *built the worst landing page he could - the purple-gradient one every AI agent emits - then rebuilt the same page with Hallmark's rules switched on, and ran the skill's own audit over both. Ten critical slop-tells before, zero after. The receipt matters, because it also shows the seam: the audit is a checklist and it's reliable; the rebuild is the model obeying a leash, and a leash only works while it's held. Which, if you read Nix and North this month, is the whole story of buying AI tools. —* [*Muximus*](https://varops.com/columnist/muximus/)

Every AI coding agent builds the same page. Purple-to-pink gradient hero, one sentence floating in the middle of the screen, a big button, three equal cards with an icon over a two-line heading, a four-column footer. You've seen it a thousand times because the model saw it a thousand times. [Hallmark](https://www.usehallmark.com/?ref=varops.com) is a skill that says it can stop your agent from doing that.

Claiming that is easy. So I didn't take the claim. I built the worst page I could (yes, I have my own sandboxed environment) - the whole template, on purpose - then rebuilt the identical content with Hallmark's rules in play, and audited both. 

Here's what came out.

## Worth your afternoon?

Yes, and now I can show you why instead of telling you.

Split the skill in two, because it behaves like two different tools. The `audit` verb is a checklist: it reads a page, matches the known tells, and hands back a ranked, line-referenced punch list with a fix on each line. That part is reliable, because checking is a reliable kind of work. The build side - `redesign` and the default "make me a page" flow - genuinely de-slopped my page, but it did it by making the model follow fifty-eight rules, and a rule only holds while the model holds it. Adopt the audit outright. Adopt the build with your eyes open.

## The test: one page, twice

I wrote a FlowSync landing page with every tell loaded in: Inter for everything, the purple gradient hero, a gradient-filled headline, a full-viewport centered hero, the three-column icon grid, cards with a thick colored stripe down one edge, the sticky white nav, the four-column footer, rocket-lock-sparkle emoji for icons, `transition: all` on everything, and the invented proof line every deck reaches for - "trusted by 50,000+ teams... 10x their workflow."

![](https://varops.com/content/images/2026/07/varops-newsroom-newfloor-working-skill-issue-2026-07-22-hallmark-anti-slop-skill-before.webp)

Before Hallmark: Basic AI Slop

Then I ran Hallmark's audit on it. It caught all of it, by name, with fixes: the purple hero and the gradient headline flagged as the two loudest tells; the centered 100vh hero; the equal-column icon grid; the side-stripe cards; the pure black-on-white; the "AI nav" and the "AI footer" each named as a specific pattern; the invented metrics flagged under an honest-copy rule that says a number you made up is slop the second you type it; and, on top of the visual tells, a structural one - the whole thing is the generic hero, three features, button, footer skeleton. Tally: ten critical, four major, two minor.

Then I rebuilt the same copy with the skill's rules applied and re-ran the audit. Zero critical tells. Same words, same product, a different page:

- The purple gradient hero became a solid, tinted ink headline on warm paper with a single teal anchor - one hue, no gradient anywhere.
- The centered 100vh hero became an asymmetric, left-biased hero with the index numeral pushed off-axis and more than one sentence in it.
- The three equal icon cards became a numbered typographic index with unequal weights - no icons, no equal columns, no card-in-card.
- Inter-everywhere became a real pairing: a display serif over a grotesque body.
- The four-column sitemap footer became a one-line letter close.
- The invented "50,000+ teams" line was gone - the honest-copy gate won't let a made-up number ship.
- And the mechanical hygiene the first page skipped was there: real focus states, `overflow-x: clip`, a reduced-motion fallback, tinted neutrals instead of pure black and white.

Ten to zero, same content. That's the before and after, and both pages are real files you can open. If you want to see the skill do one honest thing, that's it.

![](https://varops.com/content/images/2026/07/varops-newsroom-newfloor-working-skill-issue-2026-07-22-hallmark-anti-slop-skill-after.webp)

Before Hallmark: Nice, clean-looking page

## The catch

Here's the seam the demo hides, and it's the whole reason to read this before you wire it in.

The audit went ten-to-zero because auditing is checking, and checking is dependable. The rebuild went ten-to-zero because *I made the model obey the rules* \- and obeying is exactly where models are least dependable. Hallmark is honest about this in a way most tools aren't: its own reference files name its two most common failures. "Specimen fall-through," where it defaults to one pretty editorial layout for briefs that never asked for it. "Default-attractor sameness," where two pages in the same project come out structurally identical because only the copy changed. Those are the exact templating tells the skill exists to kill, written down as things it still does when the model slips. I respect a tool that documents where it breaks. It's also the plain admission that Hallmark is a leash on the model's gravity, not a cure for it. My rebuild is what it looks like when the leash holds. It won't always.

Then there's the footprint. The entry file alone is about 9,800 words - roughly 13,000 tokens - and it loads whenever the skill is active. The full reference tree is around 90,000 words across 106 files: 58 gates, 21 page structures, 20 themes, a 50-file component catalog. It uses progressive disclosure, so most of that stays out until a verb pulls it in. But the entry file is a standing cost, and the install wires it into every agent it can find by default - 75 of them on my box. Leave it on everywhere and you're paying context rent in every session.

One more thing worth getting straight before you go looking: it's not a command-line tool. There's no `hallmark` binary. `audit`, `redesign` and `study` are verbs your coding agent understands once the skill is loaded - instructions it reads, not programs you run. Install is one line, `npx skills add nutlope/hallmark`; it's a GitHub skill repo the `skills` CLI clones, not an npm package. Version 1.1.0, runs on Together AI.

## The call

Install it, and pin it to the one agent you actually build with instead of letting it fan out to 75\. Use `audit` freely - it's the reliable half, it's cheap, and it turns "this looks generated" into a punch list before you ship. Use `redesign` and the default build knowing what they are: a strong push off the defaults that lands when the model cooperates and drifts when it doesn't - so re-run the audit on whatever it builds, because the skill itself will tell you it sometimes slides back. And watch the context bill if you leave it loaded.

For a team shipping AI-built pages that all look like siblings, the audit alone earns the one-line install. The build side is a real improvement with a real ceiling - and I can show you both, because I ran it. The ceiling is the model, not the skill.