Landing page audit, implementation patch, and final OG

Use hallmark for judgment, kill-ai-slop for mechanical recall, and visual-skills for social-card direction, then deliver an inspectable page patch and final OG image.

The path and every handoff

1. 1
   Audit with judgment hallmark

   FUNCTION VERIFIED

   Find hierarchy, genre, typography, and macrostructure problems that mechanical rules cannot judge.

   Input

   Rendered page plus its audience, purpose, and declared design genre

   Instruction

   Use hallmark's audit verb and record each finding as what, where, severity, and proposed fix without editing the page yet.

   Output

   hallmark-audit.md human confirmation

2. 2
   Scan and confirm anti-patterns kill-ai-slop

   FUNCTION VERIFIED

   Catch mechanical regressions quickly, then separate true problems from design-system exemptions and scanner self-matches.

   Input

   Page source and declared design-system exemptions

   Instruction

   Run the bundled scanner, retain raw hits, and complete its required human-confirm step before proposing fixes.

   Output

   anti-slop-report.md human confirmation

3. 3
   Create the social card direction visual-skills (video + image)

   FUNCTION VERIFIED

   Choose an image model based on the content and convert the approved page direction into a constrained OG prompt.

   Input

   Approved brand direction, exact copy, palette, and 1200×630 use case

   Instruction

   Follow visual-skills' mandatory model-routing reference, freeze the exact copy and safe margins, then turn the approved direction into a final vector-first 1200×630 asset.

   Output

   og-image-direction.md human confirmation

4. 4
   Verify the corrected page and OG asset Verification or human checkpoint

   WORKFLOW VERIFIED

   Check that fixes did not introduce regressions and that the final image meets its dimensions and copy contract.

   Input

   Corrected page and generated OG image

   Instruction

   Re-run mechanical and structural checks, confirm the analytics and privacy scripts are byte-preserved, and inspect OG dimensions, text, and safe margins.

   Output

   before-after report and final-og.png human confirmation

Result you can inspect now

EVIDENCE REPLAY Nine independent landing-page delivery runs with final OG assets

2026-08-02

Open the run record — what it proves, what it doesn't, and the raw artifacts

What this proves: The same production homepage was run three times with no skill context, Hallmark alone, and the published three-skill combination. All nine runs produced a revised full HTML page, substantive audit, and renderable 1200 by 630 OG; all nine preserved the analytics and goal-privacy scripts byte for byte and passed the same structural delivery checks.

What it does not prove: That the skill combination is higher quality than the control, that arbitrary customer frameworks need no rework, or that a paid order can always be fulfilled inside three business days. Scanner medians were identical across arms.

• Valid complete deliveries: 9 of 9
• Median duration: control 126.30s, single skill 229.87s, combination 168.14s
• Median scanner result in every arm: 4 groups and 26 raw hits
• Selected final asset: combination run 2, chosen for clearest product-path communication rather than a claimed arm-level quality win

Nine-run mechanical summary →

Known limits

• The nine-run reproduction used one frozen static homepage, not nine different customer stacks.
• The service applies a patch only to supplied static source; it does not deploy to production.
• hallmark's genre severity is ambiguous when a page has no design stamp.
• kill-ai-slop's raw scanner hits require human confirmation, and scanner counts did not distinguish the three experiment arms.

Implementation pack

The pack is generated from this path record and the tested-skill source data. Install commands, commits, evidence status, handoffs, and limitations stay attached to the plan.