Repository navigation
Conversation
Any made-up /gh/<owner>/<repo>/<skill> URL answered 200 with a thin page and a canonical to the homepage. Google filed 2,621 of them as soft 404s, and the open URL space wastes a small crawl budget. A missing Skill now answers 404 with noindex and no canonical. A page with no Skill never canonicalises to the homepage.
Experiment C tests whether Skill pages that carry data GitHub lacks get crawled sooner. One constant holds both groups and the cull path.
Treatment pages swap the full SKILL.md body for facts, an excerpt, and related Skills, all from data skilld already holds. Controls stay as is.
The excerpt showed a heading and no body, and repeated the description already in the page header. It now keeps whole blocks up to about 250 words.
🤖 BLOCKED
Selected findings run after merge and open separate pull requests. |
VISION says every Skill surface leads with the run command. The reason text that pushed install for Skills with code is gone.
🤖 READY · 95/100
|
`nuxi preview --port 5678` starts wrangler on 8787 for the Cloudflare preset, so Playwright timed out waiting for 5678. The Worker also read an empty D1 under .output, so every Skill API call failed with 500 and a missing Skill page rendered its error state with status 200. The script now applies local migrations and runs wrangler dev against the same local D1 that `pnpm dev` uses. A fresh checkout gets a migrated, empty D1.
The page sets no canonical for a missing Skill, but nuxt-seo-utils adds a self canonical to every page that is not a Nuxt error. That is harmless on a 404. The spec now fails only when a canonical names another URL, which is the homepage bug the PR fixes.
A transient API error rendered a real Skill page as 200 noindex, which could drop it from Google. The failed state now answers 503 with Retry-After and emits no robots directive, so Google retries.
Neither Skill sat on a trending board after #322, so both went noindex,follow and no longer matched their indexable partners.
There was a problem hiding this comment.
Harlan Agent Kit posted this automated review. AI open source policy.
Automated Review: BLOCKED.


Stacked on #317. Merge #317 first, then retarget this pull request to
main.❓ Type of change
📚 Description
Google holds the curated Skill URLs as "Discovered, currently not indexed", and about 72% of a Skill page's text is the GitHub SKILL.md word for word. This is experiment C from the SEO death-zone note: do pages that carry data GitHub lacks get crawled sooner than unchanged pages?
Ten treatment pages swap the full SKILL.md body for an excerpt plus facts skilld already holds: run command with the install command beneath it as the opt-in, file inventory, star trend, update pace, related Skills, and compatibility or collection data when it exists. Ten star-matched control pages stay as they are. Every visitor gets the same page. No new prose, no AI text.
Both groups come from the union of the week, month, and all-time trending lists, in the production skills sitemap. The gate and both lists sit in
skill-unique-value-experiment.ts, dated, with the 2026-11-11 end date and the cull path. Delete that file and the template to end it.Treatment: obra/superpowers/writing-plans, imbad0202/academic-research-skills/academic-pipeline, emilkowalski/skills/emil-design-eng, vercel-labs/agent-skills/react-best-practices, ibelick/ui-skills/create-design-md, remotion-dev/skills/remotion-best-practices, dimillian/skills/bug-hunt-swarm, ferdinandobons/startup-skill/startup-pitch, neondatabase/agent-skills/neon-postgres, clerk/skills/clerk-orgs.
Control: mattpocock/skills/ask-matt, coreyhaines31/marketingskills/ai-seo, kepano/obsidian-skills/knap, othmanadi/planning-with-files/planning-with-files, microsoft/playwright-cli/playwright-cli, antfu/skills/nitro, addyosmani/web-quality-skills/accessibility, getsentry/skills/prompt-optimizer, browser-use/plugins/browser-use, prisma/skills/prisma-database-setup.
Open points: