clawhub-skill-packager（ClawHub 技能 packager）

ClawHub Skill Packager

Use this skill when the user wants to create, repair, review, rename, repackage, or republish a ClawHub / OpenClaw skill bundle.

Product promise

This skill makes the publish-ready skill package first.

Then it provides the separate review file that explains:

• what was provided
• what was inferred
• what was fixed
• what still deserves human review

The package is the main product. The review file is the supporting layer.

Unified identity rule

This skill uses one unified identity across runtime, packaging, and publishing surfaces:

• runtime name: clawhub-skill-packager
• slug: clawhub-skill-packager
• folder name: clawhub-skill-packager
• metadata.openclaw.skillKey: clawhub-skill-packager

Do not intentionally split runtime, folder, slug, and skill key identities unless the user explicitly asks for it.

Core job

This skill turns user input, existing skill files, or partial drafts into a publish-ready ClawHub/OpenClaw skill bundle.

It is designed to:

• inspect what is present
• infer what is missing when reasonable
• repair inconsistencies
• build the package anyway when a safe best-effort package is possible
• self-audit the result
• return one pure publish bundle plus one separate plain-text review file

Operating stance

This skill is designed for low-friction handoff.

When the user provides material:

• inspect what is there
• infer what is missing when reasonable
• choose the best safe course based on current knowledge
• avoid unnecessary clarification loops
• return a concrete package plus a review statement

Prefer statements over questions.

If something is missing but inferable:

• infer it
• note the inference
• keep moving

If something is risky, ambiguous, or likely to affect publishing:

• still produce the package when a safe best-effort package is possible
• highlight the issue clearly in the review file
• mark it for user review

Do not stop at “more info needed” when a reasonable package can still be built.

Default to the full release-bundle standard, not the minimal valid skill standard, unless the user explicitly asks for a minimal package.

Exact output contract

Always produce exactly two user-facing deliverables.

A. Publish bundle zip

A zip-ready skill folder containing only files that directly belong to the skill as a release artifact.

This bundle must include:

• SKILL.md
• README.md
• CHANGELOG.md

It may also include only those additional files that are genuinely part of the skill itself or required for the skill to function correctly, such as:

• references/
• examples/
• scripts/
• configs/
• assets/

Do not include inside the publish bundle:

• review notes
• inference notes
• build commentary
• user/AI discussion notes
• release review records
• handoff notes that are only about packaging this build

The publish bundle should look like it was made for the skill itself, not for the conversation that created it.

B. Separate plain-text review file

A separate review file in plain text.

Preferred format:

• .txt

This review file should say:

• what inputs were provided
• what information was missing
• what assumptions were made
• what was added
• what was edited
• what was removed
• what was inferred
• what still deserves human review
• whether the package appears publish-ready
• the final publish/install handoff details

The review file must remain outside the publish bundle unless the user explicitly asks for it to be embedded.

Standard support artifacts inside this packager skill

This packager skill ships with reusable support files that are part of the skill itself.

Use them like this:

• REVIEW-CHECKLIST.txt = the permanent self-audit standard
• REVIEW-RECORD-TEMPLATE.txt = the base for the generated separate review file

Operating modes

Use one of these modes based on the user's request and the material provided.

1. Package-from-scratch

Use when the user provides a concept, rough notes, or minimal package material.

Pass 1:

• assemble identity
• infer missing metadata
• define package surfaces
• identify major assumptions

Pass 2:

• build files
• self-audit
• generate the separate review file

2. Repair-existing-skill

Use when the user provides an existing skill package that needs cleanup, fixes, or modernization.

Pass 1:

• inspect existing files
• identify drift, breakage, or parser-risky structure
• determine what should be preserved

Pass 2:

• repair and normalize
• self-audit
• generate the separate review file

3. Audit-only

Use when the user wants analysis and recommendations without generating a new package.

Pass 1:

• inspect the provided package or concept
• identify strengths, issues, risks, and missing information

Pass 2:

• generate the review file
• do not build a replacement package unless the user also asked for packaging

4. Republish / update

Use when the user already has a package and wants version, naming, positioning, or packaging updates for republishing.

Pass 1:

• inspect the current package
• identify what changed since the prior release
• align versioning, naming, and publish surfaces

Pass 2:

• update files
• self-audit
• generate the separate review file

5. Rename / rebrand

Use when identity surfaces need changing while preserving intended behavior.

Pass 1:

• inspect current identity surfaces
• preserve intended behavior
• determine which naming surfaces change and which remain stable

Pass 2:

• rewrite aligned identity surfaces
• self-audit
• generate the separate review file

Packaging workflow

Step 1 — Inspect

Look at all provided material:

• text prompts
• existing SKILL.md
• existing README.md
• existing CHANGELOG.md
• file names
• intended display name
• intended slug
• invocation ideas
• notes about platform behavior
• prior package artifacts if relevant

Step 2 — Determine completeness

Check whether the package has enough information for:

• display name
• slug
• runtime identity
• version
• description
• tags
• invocation behavior
• public positioning
• file set
• frontmatter format
• compatibility with current OpenClaw parser expectations

Step 3 — Infer or repair

If something is missing or inconsistent:

• choose the best safe default
• repair mismatched naming
• align display name, slug, runtime name, folder name, skill key, and README publish fields unless the user explicitly requests a split
• correct parser-risky frontmatter
• normalize versioning
• remove obvious contradictions
• preserve intended behavior whenever possible

Step 4 — Build the package

Create the full package folder and final file contents.

Step 5 — Self-audit

Run a second-pass review using REVIEW-CHECKLIST.txt.

Step 6 — Deliver

Deliver exactly two user-facing artifacts:

• the final publish bundle zip
• the separate review file built from REVIEW-RECORD-TEMPLATE.txt

Then provide a short summary in the response telling the user:

• here is your publish bundle
• here is your review file
• here is what was inferred
• here is what was fixed
• here is what may still deserve review

Visibility rule

Do not bury final deliverables in an internal-only path without clearly surfacing them.

The publish zip and the separate review file must be:

• clearly identified
• surfaced directly to the user
• returned as the final two artifacts

Do not report completion until both artifacts are available in a directly user-visible way for the current environment.

Skill type awareness

When packaging a skill, classify it before building.

Possible classes include:

• instruction-only skill
• formatting / style skill
• workflow / orchestration skill
• code or script-backed skill
• API-dependent skill
• environment-variable-dependent skill
• binary / external-tool-dependent skill
• mixed package

Use the class to decide:

• what files are needed
• what install notes are needed
• what security notes are needed
• what review flags matter most

Runtime and security declarations

Inspect for:

• environment variable requirements
• secrets or credentials
• external API dependencies
• script or binary execution assumptions
• file system expectations
• privilege or escalation requirements
• background or always-on behavior

If these exist:

• package the skill anyway when safe
• state them clearly in the review file
• highlight them for review

Frontmatter rules

For OpenClaw compatibility, prefer:

• single-line frontmatter keys
• metadata as a single-line JSON object
• quoted version
• quoted long description strings when helpful

Preferred base frontmatter pattern:

---
name: skill-slug
description: "Short clear skill description."
version: "1.0.0"
user-invocable: true
disable-model-invocation: true
metadata: {"openclaw":{"emoji":"📦","skillKey":"skill-slug"}}
---

Naming rules

Align these surfaces unless the user explicitly wants them different:

• public display name
• package folder name
• slug
• runtime name
• metadata.openclaw.skillKey
• README publish fields
• changelog package identity

Use:

• human-readable title for display name
• lowercase hyphenated form for slug
• the same lowercase hyphenated form for runtime name by default

Review checklist

A package should be checked for all of the following.

Identity alignment

• display name matches intent
• slug is lowercase and hyphenated
• folder name matches slug
• runtime name matches slug unless intentionally split
• skill key matches slug
• README publish fields match final identity
• changelog reflects the current version and identity

Frontmatter health

• SKILL.md exists
• frontmatter is present
• name exists
• description exists
• version exists
• metadata uses single-line JSON
• invocation flags are present when useful
• emoji is present if desired
• parser-risky nested metadata is repaired

Behavioral clarity

• purpose is clear
• scope is clear
• activation behavior is clear
• explicit invocation is defined if needed
• output behavior is described
• risky ambiguity is reduced

Public positioning

• branding is reasonable
• wording is accurate
• descriptions do not overclaim capabilities
• external affiliation wording is safe when relevant

Runtime / security awareness

• skill type is correctly classified
• env var requirements are documented
• API dependencies are documented
• binaries / scripts are documented
• privilege assumptions are documented
• risky surfaces are highlighted

Deliverable separation

• publish bundle contains only skill-release files
• review content is kept outside the publish bundle
• review file is plain text
• final deliverables are surfaced separately

Deliverables

• package files were generated
• separate review file was generated
• changes are summarized
• assumptions are highlighted
• publish-readiness is stated

Severity markers

Use these markers consistently:

• ✅ FIXED AUTOMATICALLY = safe automatic repair completed
• 🔶 INFERRED FIELD = best-effort inferred value that should remain visible
• ⚠️ REQUIRED REVIEW = likely publish-affecting issue that deserves human confirmation
• 📝 EDITED FOR ALIGNMENT = consistency edit across identity or package surfaces
• 🚀 READY TO PUBLISH = no major blocker detected in the final package

Final response contract

At completion, report in this order:

1. brief status line
2. the publish bundle zip
3. the separate review file
4. short bullet summary of:

- what was created - what was changed - what assumptions were made - what should be reviewed

1. publish-readiness statement

Second-pass workflow

If the user returns with edits or clarifications:

• re-run the same inspection and package workflow
• preserve the accepted identity and structure unless the new instructions change them
• reduce the number of inferred fields
• keep the second-pass output cleaner and closer to final
• aim for a near-zero-friction publish handoff

Operating note

This skill is a packager and self-auditor for ClawHub / OpenClaw skills. Its job is to turn incomplete or inconsistent skill drafts into coherent publish-ready bundles while preserving the user's intended behavior whenever possible and minimizing decision friction for the user.