[harlan-zw]

🔗 Linked issue

❓ Type of change

• 📖 Documentation
• 🐞 Bug fix
• 👌 Enhancement
• ✨ New feature
• 🧹 Chore
• ⚠️ Breaking change

📚 Description

Maintainers need a concrete way to create one Skill for Claude Code, Codex and Gemini CLI. Add a complete example, documented discovery paths, and separate checks for selection and task completion. Lead with the project-authoring Skill's run command.

Link the guide from the homepage's authoring section, the three covered Agent pages, and project and package review steps. Keep each page's search purpose and indexability rules. Qualify the linked Agent pages' compatibility claims to match the guide.

Admit one page for how to create claude skills, measured at 390 estimated US searches/month and KD 30. Keep its source evidence outside published collections, with the existing admission removal as its cull path.

🤖 AI disclosure: Harlan Agent Kit modified this description. My AI open-source policy.

[harlan-zw]

🤖 Codex checked the built page at desktop and mobile widths.

The canonical points to /learn/create-agent-skills. The page emits index,follow, appears in the pages sitemap, and excludes editorial records. Client navigation from Learn and the project guide reaches the article and review anchor. The mobile table scrolls within its container; the document has no horizontal overflow.

The published Skill example parses through the existing loader. This does not establish live execution in Claude Code, Codex or Gemini CLI.

Desktop	Mobile

[harlan-github-agent]

🤖 MERGED

Harlan Agent Kit posted this automated review. It is not Harlan's personal review or approval. AI open source policy. Last updated: 2026-10-05 04:57 UTC.

GitHub merged this pull request.

• Logged (30/100): The PR's two rendered-layout screenshots are dead GitHub-hosted links, so the visual evidence behind the layout claims cannot be verified. View code
  • Ask an agent to verify and fix this finding

90d63c3e-33d5-4ed2-99c2-4ba75d912bc2

Selected findings run after merge and open separate pull requests.

[harlan-zw]

🤖 Harlan Agent Kit Agent assessment of head b2ad273.

An independent read-only agent reviewed the full diff, renderer and admission consumers, source claims, author screenshots, and exact-head checks. Findings: none. GitHub Actions build, lint, typecheck, tests, and harness-proof passed. Local published CLI grammar also passed.

The capacity fallback applied because four Review tasks were queued with two free host slots. The exact-head Service Review returned Stopped before independent review. This assessment does not change the Service Review status.

Additional manual evidence: Codex 0.160.0 loaded the article's exact example for explicit and unnamed matching requests, corrected the timeout claim, and asked for the missing diff. An unrelated CSS request did not load the Skill. Each used a fresh read-only session. Claude Code 2.1.288 discovered the Skill but its weekly limit blocked execution. Gemini CLI 0.54.0 rejected this client's account tier before execution. These trials prove only the observed local Codex behavior, not cross-Agent completion or generation by the authoring Skills.

Independent reviewer confidence: 90/100. Production deployment and Claude/Gemini completion remain unverified.

[harlan-zw]

🤖 Codex content-refresh review of head 47de223.

Two independent editorial reviewers rechecked the final article digest 10c533da43a4196e022b2dfe174344088e713213ca1d0237ed300880d17ddd62. The review corrected fresh-session isolation, the unsupported runtime-default inference, Gemini activation consent, description limits, and permission-configuration wording. The article's Skill body is unchanged.

The unpublished REPLAY record now preserves complete prompts, commands, versions, configured models, sanitized activation evidence, and the observed Codex outcomes. It preserves Claude discovery-only evidence and Gemini's account failure. It does not claim cross-Agent task completion.

Checked the rebuilt Worker manually at desktop and mobile widths. The corrected instructions render without document overflow. The Learn introduction now renders SKILL.md as code without a filename-domain link. Client navigation reached the article heading. The canonical remains https://skilld.dev/learn/create-agent-skills. The pages sitemap includes the article, excludes editorial records, and the replay-record route returns 404.

Production delivery and Claude/Gemini task completion require separate observation. This editorial assessment does not change the Service Review status.

[harlan-zw]

🤖 Harlan Agent Kit Agent assessment of head 47de223.

Premise: sound. Findings: none. An independent read-only agent reviewed the complete net diff, affected renderer and admission consumers, source-backed consent wording, final article digest, durable replay records, and latest desktop/mobile captures. The inline-code correction removes the false filename link without changing article claims.

The capacity fallback applied because three Review tasks were queued with two free host slots. The Service returned Stopped for this exact head before the independent assessment. This assessment does not change the Service Review status.

The article's exact digest has separate independent factual and editorial acceptance recorded in the brief. Claude/Gemini task completion and production delivery still require observation. GitHub Actions checks were pending during assessment; the author retains the CI gate.

Independent reviewer confidence: 95/100.

[harlan-zw]

🤖 Codex internal-linking follow-up for head bb94d26.

The built Worker returned server-rendered links from eleven source pages. Homepage, Claude Code and Codex are the three indexable donors. Gemini, Learn and project/package guides retain their noindex rules. No admission or canonical rules changed.

All nine new links reached the article through client navigation. The five package-guide links reached the existing review fragment. Desktop/mobile authoring placement showed no clipping, overlap or document overflow. The strategy record returned 404 and stayed outside the sitemap.

The independent editorial review accepted contextual anchors and the donor qualifications. Retained NuxtSEO evidence could not establish a donor with proven click traffic or a ranking lift. The unpublished link map records those limits and dated-window measurement instructions.

[harlan-github-agent]

Harlan Agent Kit posted this automated review. AI open source policy.

Automated Review: READY.

Read the Review results