SkillHub打包、关键词与搜索优化

概述

SkillHub Publish Optimizer

Prepare an installable, trustworthy SkillHub release. Preserve the skill's actual behavior; optimize how users and agents find and understand it.

Workflow

1. Inspect the skill's SKILL.md, scripts, and declared dependencies. State the supported platform, required credentials, and meaningful user actions before drafting metadata.
2. Audit the source tree. Do not package credentials, private keys, .env files, user data, build caches, Git metadata, or previous archives.
3. Research the market using [references/skillhub-search.md](references/skillhub-search.md). Extract 2-4 Chinese and English query variants from the task, object, audience, platform, and technology. Search every candidate with sortBy=score; inspect the top results' names, summaries, categories, and popularity.
4. Draft one primary listing and, when useful, one alternative name. Keep terminology consistent with the actual capability and search evidence.
5. Create or update a release-side publish-form.md containing display name, summary, category, version, changelog, pricing, requirements, and explicit limitations.
6. Run scripts/release-skill.ps1 to sync the skill source into its Git repository, validate, package, commit, and optionally push in one deterministic operation. The archive must contain SKILL.md at its root.
7. Create a GitHub repository through the authenticated GitHub UI only when no suitable remote exists; then let the release script add origin and push main.
8. In the authenticated SkillHub publishing UI, upload the ZIP, fill the release form from publish-form.md, select the matching categories, and submit. Never enter passwords, SMS codes, or scan a QR code on the user's behalf. Pause visibly at login and resume only after the user confirms authentication.
9. Poll the team dashboard every 10 seconds only while the release is in 安全审核中. When it changes to 团队管理员审核中, switch to the administrator review page, verify slug/name/version/changelog, confirm approval, and verify the final 已发布 state. Stop and report if the release is rejected or remains unchanged for the agreed timeout.

Listing Copy

Use these rules for the display name and summary:

• Put the primary task or object near the beginning of the display name. Add the platform or technology only when it differentiates the skill.
• In the first sentence, state what the skill does, who or what uses it, and the principal result. Follow with concrete supported actions and relevant technical terms.
• Cover search variants naturally: Chinese and English terms, common abbreviations, and one or two task synonyms. Prefer terms users would genuinely search.
• Treat keyword lists as research notes, not publish metadata. SkillHub's current publishing form has no editable tag field; place only the 2-4 most relevant terms naturally in the name and first sentence of the summary.
• Select the category for the primary user job, then a specific available subcategory. Do not choose a category merely because a broad keyword is popular.
• State OS limits, required API keys, pricing, local/cloud processing, and any security boundaries. This reduces unsuitable installs and improves user experience.

Do not promise unsupported operations, claim a ranking improvement, copy competitors' wording, or stuff repeated keywords. Ranking formula details are not public; relevance, classification, source, and user adoption all matter.

Release Form

Use this compact format in publish-form.md:

# SkillHub Publish Information

- Slug: `<slug>`
- Display name: `<task-first name>`
- Summary: `<what + audience + key actions + differentiator>`
- Category: `<primary category>` (subcategory: `<subcategory>`)
- Keyword research: `<2-4 search terms used for baseline and retest; not a publish-form field>`
- Version: `<semver>`
- Changelog: `<user-visible change>`
- Pricing: `<free/paid>`
- Requirements and limits: `<OS, API keys, local/cloud, exclusions>`

Packaging

Run from PowerShell after the copy and version are final:

.scriptspackage-skillhub-release.ps1 -SkillPath <skill-folder> -OutputZip <release.zip>

Use -ListOnly to review selected files. Existing output files require -Force to overwrite. A secrets directory is excluded only when it contains no data other than .gitkeep; any credential-like file stops the release.

Deterministic Release Command

Use the helper below rather than manually repeating the validation and archive checks:

.scriptsprepare-skill-release.ps1 -SkillPath <skill-folder> -OutputZip <release.zip> -Force

It runs the official basic validator, invokes the safe packager, and verifies that the ZIP contains root-level SKILL.md. It never uploads, logs in, creates a repository, reads credentials, or changes the source folder.

For the full local-to-GitHub handoff:

.scriptsrelease-skill.ps1 -SkillPath <skill-folder> -RepositoryPath <git-folder> `
  -OutputZip <release.zip> -CommitMessage "Release <version>" -Push

The script synchronizes only skill-owned files (SKILL.md, agents/, references/, scripts/, publish-form.md, and CHANGELOG.md), leaving repository README/license files intact. It refuses a non-Git directory, never stages secrets or archives, and never handles credentials. Use -NoCommit or -NoPush for review-only runs.

Browser Handoff

Use the published publish-form.md as the single source for slug, display name, description, version, and changelog. Before clicking publish, compare every field against that file and verify the selected ZIP file count/size. Treat the final publish click and administrator approval as explicit actions authorized only when the user asks to publish or approve. Keep the login page visible and concise: state that the user needs to authenticate, then continue from the existing tab.

Verification

• Run C:UsersAdministrator.DESKTOP-0KF8V7V.codexskills.systemskill-creatorscriptsquick_validate.py <skill-folder>.
• Confirm the ZIP contains root-level SKILL.md, agents/, scripts, and only required templates or references.
• Record keyword, category, date, search sort, result count, rank, and the five nearest competitors. Treat rank changes as observations, not proof of causality.