<- all tokdocs
2026.10.24@thehypekevin

Four design skills, four real repos, two things the video gets wrong: Hallmark's theme is "Piste" not "Pieced," and Impeccable does run on its own on Codex once you approve its hook

AI
Watch on TikTok

View on TikTok ->

The artifact is a 407.80 s vertical MP4 (mov,mp4,m4a,3gp,3g2,mj2 container, 6,432,117 bytes, format id h264_540p_126180-1, container bitrate 126,180 bps) holding one H.264 High-profile video stream at 576x1024, yuv420p, 30/1 fps, 12,232 frames, 407.73 s, 87,736 bps, and one AAC HE-AACv2 stream at 44,100 Hz stereo, 32,009 bps, 407.80 s; ffprobe and metadata.json agree on every field. It was posted at epoch 1792861200, which is 2026-10-24 17:00:00 UTC, a Saturday, by thehypekevin (nickname "The Hype Kevin", uploader id 6894383934525588481), credited to "original sound - The Hype Kevin," with a 392-character description that names the four skills and the voting site. At capture time, the mtime of metadata.json, 2026-10-26 05:22:17 UTC, the post showed 2,441 views, 71 likes, 2 comments, 8 reposts and 48 saves. I read 22 of the 204 extracted frames: 002, 012, 020, 022, 030, 043, 050, 061, 068, 074, 090, 110, 133, 146, 171, 181, 183, 184, 185, 188, 198 and 204; the Whisper transcript has 112 segments and 1,388 words. Frame 002 (2-4 s) opens on a lower third reading "MADE-UP EXAMPLE · Rally · Dink District" above a neon-yellow landing page headlined "Every court. Every hour. One board." Frame 012 (22-24 s) shows a 2x5 grid under the column heads "Claude Code" and "Codex," with a chip reading "Why Every AI App Looks the Same" and a pickleball mascot asking "Which skill changes what your agent builds?" Frame 020 adds label chips including "Taste Skill" and "UI UX Pro Max." From frame 022 onward every build appears as a browser mock whose address bar reads rally.localhost, captioned with burned-in run labels: "Claude Code · plain · r1," "Claude Code · UI UX Pro Max · r1" (frame 030), "Claude Code · Hallmark · r1" over a foggy-valley hero reading "Pickleball prime time, without the clipboard." (frame 043), "Claude Code · Taste · r1" over a yellow "Rally" masthead (frame 050), "Claude Code · Taste · r2" over an aerial court photo tagged per court ("Court 3 · Kowalski, doubles") with time chips "5:00 pm / 6:00 pm / 7:00 pm" (frame 061), "Claude Code · Impeccable · r1" and "· r2" (frames 068 and 074, the second a brown signpost labelled "Court bookings," "Open play," "Leagues and ladders PRO," "Clinics PRO," "Waitlist," "Check-ins and payments"), a red-ringed testimonial tagged "placeholder" reading "Maya Okonkwo, General manager, Dink District · 8 courts" beside pricing of "Club $29 / Pro $79 Most popular / Network $199" (frame 090), and "Codex · Hallmark · r2 · Piste" over "MORE PLAY. LESS RUNAROUND." (frame 110). Frame 133 shows the plain Claude clash card, "Double booking · 6:00 pm / Court 3 is booked twice," "in 3h 20m," pitting "Henderson doubles / Carol Henderson + 3 · 6:00-7:30 pm · Booked Oct 14 in the app · paid $48" against "Ladder round 6 / Priya Shah vs Dev Malhotra," with the option "Move the ladder match to Court 6, same time." Frame 146 stacks a left-rail "plain · r1" dashboard against a top-nav "Hallmark · r2" court sheet. Frame 181 is a bar chart headed "What does it cost?" reading "Claude · Taste 41 min" and "Claude · plain about 19 min." Frames 184 and 185 are the only frames carrying install commands, and frames 188 and 204 show the arena.thehypekevin.com comparison asking "which would you ship?" The spoken argument is that four installed design skills change what Claude Code builds in three cases out of four, that Codex overrides most of what a skill tells it, and that redesigning an existing app with any of them repaints rather than relayouts. On the transcript-end check: the audio runs 407.80 s, segment 111 is timestamped 402.96 to 412.88 s and so begins inside the audio but runs past it, and its text, "Of codex.com.," does not match frame 204 at 406-408 s, which shows the arena comparison and no such URL; segment 112, "You" at 421.54 to 427.46 s, lies wholly beyond the audio and would need a frame 211 in a 204-frame extraction, so it is fabricated. Neither tail is quoted below.

The install card is the only hard evidence in the video, and all four commands resolve

The video never shows a terminal, an agent session, a token meter or a stopwatch. The single frame that carries verifiable identifiers is frame 184, held through frame 185 (366-370 s), which lists four commands verbatim:

Hallmark        npx skills add nutlope/hallmark -g
Impeccable      npx impeccable install
Taste           npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"
UI UX Pro Max   npx ui-ux-pro-max-cli init --ai claude

Every one of those resolves to a live primary source. npx skills add is the Vercel Labs skills CLI, whose README documents both the -g flag ("project by default; add -g to make them available across projects") and the --skill selector used in the Taste line. github.com/Nutlope/hallmark exists, created 2026-04-27, described as "Anti-AI-slop design skill for Claude Code, Cursor, and Codex." github.com/Leonxlnx/taste-skill exists, created 2026-02-19; the capitalisation in the on-screen URL is correct, and the skill name in the --skill flag matches the name: design-taste-frontend field in the frontmatter of skills/taste-skill/SKILL.md. The npm package impeccable is real, published from github.com/pbakaus/impeccable, latest version 4.7.0, last modified 2026-10-19. The npm package ui-ux-pro-max-cli is real at version 2.15.0, and its README documents uipro init --ai claude alongside --ai codex, --ai cursor and nine more targets. Star counts read from the GitHub API on 2026-10-26: UI UX Pro Max 140,631, Taste 102,721, Impeccable 88,399, Hallmark 34,971.

"A theme called Pieced" is a transcription error; the theme is Piste, and it ships in the repo

At 216.42 s the video says "And Hallmark's second run dealt a theme called Pieced." Whisper heard "Pieced." The burned-in label in frame 110 (218-220 s) reads "Codex · Hallmark · r2 · Piste," and the on-screen caption at the same moment reads "theme called Piste,". The repository settles it: skills/hallmark/references/themes/ contains piste.md with a matching piste.jpg. No file named anything like "pieced" exists in that directory. This is the same failure mode the TokDoc corpus has logged before, where a product name is flattened into a plausible English word; anyone acting on the spoken audio would search Hallmark's theme list for a theme that does not exist.

The theme count in the same passage holds up. At 83.58 s the video says Hallmark "starts from what the visitor came to do, then deals one of 48 themes." The themes directory contains 49 Markdown files, one of which is README.md, leaving 48 themes, and exactly 48 matching .jpg previews. The job framing is also in the source: SKILL.md step 3 reads "Direction: deal a theme, or design freely when asked for custom," and the draw script is invoked as theme.mjs draw --material photo|product|words --job <job> --theme <slug>.

The "big 2.0 release last Thursday" lands on the right day

At 83.58 s the video says Hallmark "had its big 2.0 release last Thursday." The post is dated 2026-10-24, a Saturday, so "last Thursday" means 2026-10-22. The frontmatter of skills/hallmark/SKILL.md carries version: 2.0.0. The commit log shows "Claude Code plugin and marketplace files at 2.0.0, from the idea in #75" on 2026-10-21T17:36:50Z, then "Merge pull request #79 from Nutlope/v2" at 2026-10-22T16:32:55Z, with the final 2.0 site commits landing at 2026-10-22T17:20:30Z. The repository's pushed_at is 2026-10-22T17:20:32Z. The claim checks out on the day. One caveat worth recording: Nutlope/hallmark publishes no GitHub Releases and carries no git tags, so there is no release artifact to point at, only the merge of the v2 branch. Anyone trying to pin "2.0" to a release page will not find one.

Pro Max really did say team red and championship gold, and the recommendation is a row in a CSV

At 61.14 s the video says Pro Max "is a big database of styles and palettes, and for your app, it recommended Team Red, Championship Gold, and a stadium font," then at 69.30 s, "And the agent built dark green and lime anyway, on all four Pro Max runs." The first half is traceable to a file. Row 75 of src/ui-ux-pro-max/data/colors.csv in nextlevelbuilder/ui-ux-pro-max-skill is the "Sports Team/Club" entry, primary #DC2626, accent #EF4444, and its notes column reads, verbatim, "Team red + championship gold [Accent adjusted from #FBBF24]." The matching product row in products.csv is also number 75, "Sports Team/Club," keyed on "club, sports, team," recommending "Vibrant & Block-based + Motion-Driven" with "Team colors + Energetic accents." The "stadium font" is looser language than the data uses: the nearest entry is pairing 49 in typography.csv, named "Sports/Fitness," which pairs Barlow Condensed for headings with Barlow for body and is tagged "sports, fitness, athletic, energetic, condensed, action." Nothing in the catalog is called a stadium font.

The scale claim is also right. The README advertises 79 searchable UI styles with 50 active, 192 color palettes aligned one-to-one with 192 product types, and 192 reasoning rules, backed by the CSV files in src/ui-ux-pro-max/data/ that I listed through the GitHub contents API. The second half of the claim, that all four Pro Max runs ignored the recommendation and shipped dark green and lime, is not checkable from outside; frame 030 shows one Pro Max run in dark green with a lime button, which is one data point consistent with the claim and not a verification of "all four."

"On Codex it doesn't run on its own" is a partial correction: it does, after you approve the hook

At 243.72 s the video says "if you use Impeccable on Codex, tell it to run its checker, because on Codex it doesn't run on its own." The Impeccable README says the opposite in its default case and then names the catch. Under "Design hook" it states that on "Claude Code, GitHub Copilot, Codex, Cursor, and Grok Build, npx impeccable install and npx impeccable update install a provider-native hook manifest along with the skill payload," and that "Claude Code, GitHub Copilot, Codex, and Cursor surface findings after the edit." The Codex entry in the same list is concrete: ".codex/hooks.json runs .agents/skills/impeccable/scripts/impeccable hook, with a commandWindows sibling that calls impeccable.cmd for cmd.exe."

The reason the video's experience differs is a trust gate, which the README calls out twice. First: "Codex users should open /hooks after install or update and approve the project hook when prompted. Codex tracks trust by hook definition, so updates that change .codex/hooks.json can require approval again." Second: "Codex requires one platform step that Impeccable cannot safely skip: open /hooks after install or update and approve the project hook. There is no Codex marketplace/plugin install flow for this hook." So the accurate statement is that Impeccable's post-edit detector does run on its own on Codex once the project hook is approved in /hooks, and does not run at all until it is. The video's advice to invoke the checker manually is a workaround for an unapproved hook rather than a property of Codex support. The related claim at 81.08 s, "Impeccable checks the code after every edit," is correct in substance: the hook fires on direct UI file edits and surfaces findings back into the agent flow, running 62 deterministic detector rules with no LLM and no API key, per the README's own summary line, "1 skill, 24 commands, live browser iteration, and 62 deterministic detector rules."

Taste's 41 minutes is unsourced, but the mechanism behind it is documented

At 356.00 s the video says "Taste is the slow one. It took 41 minutes on Claude, where Plain took about 19, because it opens the page in a real browser and critiques its own screenshots." Frame 181 (360-362 s) renders this as a two-bar chart labelled "Claude · Taste 41 min" and "Claude · plain about 19 min." Those two numbers have no visible provenance anywhere in the video. No terminal, no log, no timer and no run count appear in any frame I read, and the video never says whether 41 minutes is one run, a mean of two, or the longer of two. Treat both figures as one creator's single-machine observation, not a benchmark.

The stated cause is documented. skills/taste-skill/SKILL.md describes the skill as "tasteskill v2" and lists four mechanisms, including "A house style with no ornaments, measured in a real browser by scripts/inspect.mjs, so the rules are checked rather than hoped for" and "You look at what you made. Screenshots at three widths, a juror's rubric, fix, look again." The dealing behaviour the video describes at 75.44 s, "Taste deals the agent a direction before it designs anything, a palette and a font, like a card," matches the SKILL.md line that scripts/draw.mjs "deals three of twenty researched directions; you choose the one that carries the subject's idea, and the script then deals its type system, palette, hero and section compositions from a library distilled from 247 real sites." The attribution at 114.12 s, "It found those on Wikimedia, with credits," matches the asset search step, which calls node <skill>/scripts/assets.mjs "<query>" against Openverse and Wikimedia. A browser loop plus three screenshot widths plus a fix pass is a plausible mechanism for roughly doubling wall-clock time, and the specific minute figures remain unverified.

What the video shows, what it does not show, and how 10 builds become 30

At 0.00 s the video says the app was "built 10 times from the exact same brief," while the title and description say "built thirty times" and the narration at 30.12 s says "30 builds in all." These reconcile rather than conflict. The grid in frame 012 holds ten cells, two agents by five setups, which is the ten the opening line counts. The run labels in the frames carry r1 and r2 suffixes for the greenfield landing pages and dashboards, and the redesign pass at 320.38 s is counted as "All 10 kept the exact same headline." Ten setups built twice for the greenfield round plus one redesign each gives thirty, which matches the arena site's own copy offering access to "all thirty builds." Nothing in the video states that arithmetic outright, so a viewer who only hears the opening line will undercount by two thirds.

The absences matter for how much weight the comparison carries. Across all 22 frames I read, there is no terminal window, no Claude Code or Codex session transcript, no model name, no cost or token readout, no timing instrumentation and no display of the brief itself beyond spoken paraphrase at 254.30 s ("court 3 is double booked at 6, the Hendersons doubles against the ladder match"). Every build is presented as a rendered browser mock at rally.localhost. Each cell of the grid is one or two runs, which is too small a sample to separate a skill's effect from run-to-run variance, and the video is honest about that in its framing: the conclusion at 394.76 s is "On Claude Code, three of the four did," which is a count of four comparisons, not a measurement. The video is also careful where it matters most: the lower third in frame 002 reads "MADE-UP EXAMPLE · Rally · Dink District" from the opening seconds, and the placeholder testimonial is flagged on screen with a red ring and a "placeholder" tag in frame 090 before the narration warns at 185.96 s, "Just don't ship her."

Key Takeaways

  • Correction: The video says at 216.42 s that Hallmark dealt "a theme called Pieced." The theme is Piste. The burned-in label in frame 110 reads "Codex · Hallmark · r2 · Piste," and Nutlope/hallmark ships skills/hallmark/references/themes/piste.md with a matching piste.jpg. No theme named Pieced exists.
  • Partial correction: The video says at 243.72 s that Impeccable's checker "on Codex it doesn't run on its own." Impeccable's README documents a native Codex hook at .codex/hooks.json that surfaces findings after each edit, gated on one manual step: "open /hooks after install or update and approve the project hook." It runs on its own once approved.
  • Verified: All four install commands in frames 184 and 185 resolve to live primary sources: Nutlope/hallmark, pbakaus/impeccable (npm impeccable 4.7.0), Leonxlnx/taste-skill (skill name design-taste-frontend, matching the SKILL.md frontmatter), and nextlevelbuilder/ui-ux-pro-max-skill (npm ui-ux-pro-max-cli 2.15.0, which documents init --ai claude).
  • Verified: Hallmark deals one of 48 themes. The themes directory holds 49 Markdown files including README.md, leaving 48, plus 48 matching previews.
  • Verified: Hallmark's 2.0 landed on the Thursday the video names. SKILL.md carries version: 2.0.0, and PR #79 from branch v2 merged at 2026-10-22T16:32:55Z, two days before the 2026-10-24 17:00:00 UTC post.
  • Verified: Pro Max's sports recommendation is a literal catalog row. colors.csv row 75, "Sports Team/Club," has the notes string "Team red + championship gold [Accent adjusted from #FBBF24]."
  • Clarification: The video calls Pro Max's recommendation "a stadium font" at 61.14 s. The catalog's closest entry is typography pairing 49, "Sports/Fitness," pairing Barlow Condensed with Barlow. Nothing in the data is named for a stadium.
  • Clarification: "UI UX Pro Max, the most starred design skill on GitHub" (51.60 s) holds among single-purpose design skills, at 140,631 stars on 2026-10-26 against Taste at 102,721, Impeccable at 88,399 and Hallmark at 34,971. anthropics/skills, the multi-skill repository that contains frontend-design, carries 183,875 stars, and its individual skills have no separate counts.
  • Clarification: Hallmark publishes no GitHub Releases and no git tags, so "2.0" is traceable to a frontmatter field and a merged branch, not to a release page.
  • Unverified: The 41-minute Taste run and the roughly 19-minute plain run (356.00 s, frame 181). No frame shows a timer, a log or a run count. Treat them as one creator's single-machine observation.
  • Unverified: "The agent built dark green and lime anyway, on all four Pro Max runs" (69.30 s). Frame 030 shows one such run. The other three are asserted, not shown in any frame I read.
  • Context: Ten setups and thirty builds are both correct. The grid in frame 012 is two agents by five setups, which is the ten the opening line counts; two greenfield runs each plus one redesign each totals thirty, matching the arena site's "all thirty builds."
  • Context: Whisper's segment 111, "Of codex.com." at 402.96 to 412.88 s, runs past the 407.80 s audio and contradicts frame 204; segment 112, "You" at 421.54 to 427.46 s, lies entirely beyond the audio. Neither is quoted here.

Resources

Published October 24, 2026. Writeup generated from a favorited TikTok.