Hallmark and Impeccable Both Promise Agents That Stop Making AI-Looking Sites. Their Issue Trackers Disagree.

Two GitHub repositories now sell the same fix for the same complaint: coding agents build web pages that look machine-made. Nutlope's Hallmark shows 33.8k stars under an MIT licence, and Paul Bakaus's Impeccable shows 87.7k under Apache 2.0, both as of Oct. 26. Both install into Claude Code, Codex and Cursor. Neither page offers independent evidence that the output looks less generated, and each repository's open issues describe a different kind of unfinished work.

What each one does

Hallmark's README tagline reads: "A design skill for Claude Code, Codex, Cursor and OpenCode that refuses to look AI-generated." The page says it was made by Together AI. It picks from 48 "material" themes, builds a page around the visitor's main action, links to 755 openly licensed photos, and ends with a linter that renders the page at eight widths before a fresh-context subagent reviews it. Checks are sorted into three tiers called Floor, Reflex and Finish.

Impeccable's tagline is "The design language that makes your AI harness better at design." The README lists 24 commands, including audit, polish, critique and animate, plus 62 deterministic rules in a detector that scans code or rendered pages without a model or API key. It writes product context to PRODUCT.md and visual decisions to DESIGN.md. The page credits Bakaus as creator and gives no job title.

	Hallmark	Impeccable
Stars (Oct. 26)	33.8k	87.7k
Licence	MIT	Apache 2.0
Open issues	20	25
Commits	297	2,562

1. Impeccable87.7 thousand
2. Hallmark33.8 thousand

Source: github.com/Nutlope/hallmark and github.com/pbakaus/impeccable, accessed 2026-10-26

Impeccable's own site shows demo figures, among them a largest-contentful-paint time falling from 3.4 seconds to 1.1 seconds, JavaScript dropping from 412 kB to 86 kB and a sample critique scoring 61 of 100. Those come from illustrative examples, not from a documented study, and its testimonials are social posts, one of them from a user who built a landing page in about two hours. Stars measure interest, not use. GitHub's trending list showed Hallmark gaining 1,317 stars and Impeccable 1,092 in a day, which says more about timing than about who ships better pages.

The idea comes from Anthropic's own plugin

Impeccable's README says it builds on Anthropic's frontend-design skill, which it calls the first widely used design skill for Claude. That plugin lives in the anthropics/claude-code repository, and its README describes it as one that "Generates distinctive, production-grade frontend interfaces that avoid generic AI aesthetics." The README lists Prithvi Rajasekaran and Alexander Bricken as authors and gives no titles.

The Anthropic version is a README and a single skill folder, with no licence file in the directory listing. The newer two are toolchains: a linter, a browser check, a theme library. That is the real difference. Anthropic's plugin tells a model what to avoid. Hallmark and Impeccable try to catch it after the fact.

What the issue trackers show

Hallmark's open issues skew toward the skill itself. Issue #76, opened Sept. 25, says relative links in SKILL.md break after npx skills add. Issue #62 reports contradictions when following the skill literally, and #50 asks for the roughly 1 MB skill to be split into several. Issue #37 notes that the slop test reports 58 of 58 passing while only 57 gates exist. Issue #74 carries a measurement in its title: Hallmark cuts pure black and white from 100 percent to 50 percent of pages. That is the reporter's number, and the repository does not say it has verified it.

Impeccable's tracker is busier and more technical. Issue #1175, opened Oct. 26, says its 16 MB engine binary is copied into every harness's skill folder. Issue #1142 says the static detector reads absolute CSS units such as pt, mm and cm as em, producing false findings on print pages. Issue #988 says values from mutually exclusive @media blocks get mixed together, which produces false low-contrast reports. Issue #1023 notes that the ChatGPT plugin still shows version 4.3.1 while the skill is at 4.5.2.

The pattern is consistent. Hallmark's reported faults are about whether the instructions agree with each other. Impeccable's are about whether its detector is right, and a detector that is wrong in a CI check blocks a pull request.

Telemetry and caveats

Impeccable's README says telemetry is on by default. Choosing a design sends the roll key, the choice data and the client version to impeccable.style. Users can turn it off with npx impeccable telemetry off, DO_NOT_TRACK=1 or IMPECCABLE_NO_TELEMETRY=1. The project's own caveats are plain: a clean detector run is not proof of visual or accessibility quality, and live mode is meant for trusted local checkouts only. It is not alone: NVIDIA's OpenShell agent sandbox also ships with telemetry on by default, so a developer adding agent tooling to a project now has a default to check each time.

Hallmark's page lists a different limit: browser checks and screenshots need puppeteer-core and Chrome, and without them only static checks run. Its photo library links to Wikimedia Commons images and excludes the major stock sites.

Both projects sit in a crowded category of agent skills that bundle prompts with scripts. Cloudflare's security audit skill is a 450-line seed, and JetBrains' ThinkRail and Corgi are competing for the place agents run. A design skill that ships a linter has to be right about CSS, which is a harder standard than being right about taste. The thing to watch is whether the false-finding bugs, #1142 and #988 among them, close before the next Impeccable release.