Skill: publish-skill

Phase 0: Init and Identify Source

Required Inputs

Collect from the user:

1. Source skill path: directory containing the personal skill (e.g., ~/.claude/skills/my-skill/ or ~/.agents/skills/my-skill/)
2. Target package path: directory of the distributable package (e.g., ~/workspace/<your-package>/)
3. Target license: license of the package — ask; offer MIT as the default, never assume it

Actions

1. Read SKILL.md from the source skill directory
2. Inventory all files recursively (ls -R)
3. Classify skill type:
   • Standalone: self-contained skill with no agent delegation
   • Orchestrator: delegates to sub-agents (NOT suitable for distribution without refactoring)
   • Wrapper: thin wrapper around another tool/API
4. Present inventory table to the user:

| File | Lines | Type | Notes |
|------|-------|------|-------|

Make every later fix on a working copy (<cleaned_skill_path> below) or in the target directory. NEVER modify the source skill in place, because it is the user's working original.

Gate: User confirms source skill and target package before proceeding.

---

Phase 0.5: Skill-Worthiness Gate

Before spending effort on PII scrubbing and generalization, confirm the workflow is worth distributing as a skill at all. Apply all three gates:

Gate	Question	Pass condition
Uniqueness	Could a competent user get the same result by searching the web for ~5 minutes, or by asking a general assistant with no skill installed?	No
Specificity	Does it encode a workflow, decision heuristic, constraint, or convention specific to this domain or a recurring task — rather than a generic code snippet or a standard-library example?	Yes
Effort	Did discovering it take real debugging, study design, operational effort, or a reviewer-anticipation lesson (a pitfall, a verification step, a domain convention)?	Yes

Gate: Any "no" (or "yes" on the inverse) stops publication — recommend documentation or a memory note instead. If the value is real but the skill delegates to private agents, route through Phase 1's orchestrator finding (refactor to standalone first). Only a clear three-way pass proceeds.

---

Phase 1: Originality Check

Checks

1. External source: Is this skill adapted from another package or author? Check for attribution headers, license blocks, or "based on" comments.
2. Third-party content: Do any files in references/ come from external sources (published guidelines, textbooks, standards bodies)?
3. Competitive sensitivity: Does the skill reveal proprietary business logic or competitive advantage that should remain private? Ask the user; never make this judgment call yourself.

Decision Matrix

Finding	Action
Fully original	Proceed to Phase 2
Adapted with compatible license	Add attribution header, proceed
Contains non-compatible third-party content	Flag for removal or URL reference conversion
Orchestrator with private agent references	STOP -- requires refactoring to standalone first
Competitive/proprietary logic	STOP -- not suitable for open-source

---

Phase 2: PII De-identification Audit

Zero tolerance: the skill must have exactly 0 PII matches before proceeding.

Pre-scan Setup

Before running, ask the user for everything unique to them that should also count as a PII hit:

• Their name(s) in all languages and romanizations (a placeholder shape: <First Last>|<native-script name>)
• Their institutional affiliation(s) (a placeholder shape: <Institution>|<Hospital>)
• Any collaborator surnames that may appear in drafts or filenames

Combine the inputs into a single grep -E alternation pattern (pipe-separated).

Automated Scan

Run the bundled audit script. The first argument is the skill directory; the second is the user-specific alternation pattern from Pre-scan Setup.

bash ${CLAUDE_SKILL_DIR}/scripts/audit_skill.sh --strict <source_skill_path> \
    "<First Last>|<native-script name>|<Institution>|<Hospital>"

The script scans ten categories, plus the user pattern:

1. Hardcoded paths (/Users/<name>/, /home/<name>/, ~/Documents, ~/Desktop, ~/Downloads, ~/Projects)
2. Email addresses (any address-shaped string)
3. IP addresses / internal URLs (*.internal, *.local, *.corp)
4. Institutional references (SNUH / AMC / SMC / KAIST / SNU / ASAN / MGH / UCSF / Mayo Clinic / Johns Hopkins / Samsung Medical / Severance / Asan Medical)
5. Academic roles with names (professor <Surname>, Prof. <Surname>, Dr. <Surname>, PGY[0-9], <한글이름> 교수님)
6. Language hardcoding ("in Korean", "한국어로", "in Japanese", "in Chinese")
7. Location specifics (Seoul / Busan / Daegu / Tokyo / Beijing / Shanghai / Boston / Stanford and Korean variants)
8. Blockquote dated precedent (> YYYY-MM-DD ... lines that reveal an internal review timeline)
9. Author-style filenames (<Surname>{Year}_* pattern, e.g., <Surname>2025_<Journal>_Fig01.png; allow-list excludes generic tokens like Issue2024_, Sample2025_)
10. Binary EXIF metadata (DOCX / PPTX / XLSX / PDF / PNG / JPG / TIFF — scanned via exiftool for home paths, email addresses and the user pattern, matched case-insensitively. If binaries are present and exiftool is not installed, the RESULT line reads INCOMPLETE and --strict exits 3)

Exit codes: 0 clean, 1 findings, 2 usage error / invalid user regex / a scan that could not run, 3 (--strict only) a check that could not run. Only exit 0 means clean.

Known limits: the Korean role branch (item 5) requires the honorific -nim suffix. A bare third-person mention (a Korean name followed by the role word without -nim) is not caught by this script, because separating it from job descriptions such as "advising professor" needs a stoplist that grep -E cannot express. Check such mentions by hand.

Cross-validation

For categories the script flags, also verify manually with the Grep tool against ${CLAUDE_SKILL_DIR}/references/pii-patterns.md. Pay particular attention to:

• Names not in the extra-patterns argument (e.g., a co-author who appeared only in one early draft)
• Domain-specific institutional acronyms (your institution may not be in the default list)
• Project-specific identifiers like CK-NN, MA-NN, dated cohort names

Output Format

Present all findings in a remediation table:

| # | File:Line | Category | Match | Suggested Fix |
|---|-----------|----------|-------|---------------|

Gate: User reviews all findings. Fix each one in the working copy. Re-run the audit on the working copy. Proceed only when 0 hits are confirmed.

---

Phase 3: Generalization

Language

• Replace: "in Korean" / "한국어로" / "Korean language" → "in the user's preferred language"
• Replace: "communicate in [specific language]" → "Communicate with the user in their preferred language"
• Keep: multilingual trigger keywords in the metadata.triggers field (or the target package's equivalent) (these aid discovery)

Role

• Replace: "radiology researcher" → "medical researcher" (if the skill is domain-general)
• Replace: "professor" / "fellow" → "researcher" or "user" (context-dependent)
• Keep: domain-specific terms that define the skill's scope (e.g., "diagnostic accuracy" is fine)

Paths

• Replace: hardcoded absolute paths → ${CLAUDE_SKILL_DIR} for bundled reference files
• Replace: ~/Documents/... → user-provided output directory
• Keep: relative paths within the skill directory structure

Environment

• Remove: assumptions about specific OS (macOS, Linux)
• Remove: assumptions about specific editors or IDEs
• Remove: references to personal infrastructure (agents, other personal skills)
• Keep: tool requirements, described in the skill body (a prerequisites section). The schema has no tools: field, and allowed-tools pre-approves tools rather than describing them, so do not move requirements there

Interoperability

• Check: does the skill reference other skills by name (e.g., "route to analyze-stats")?
• If referenced skill exists in target package: keep the reference
• If referenced skill does NOT exist in target package: make it optional with fallback instructions

Output

Show a unified diff of all generalization changes for user review.

---

Phase 4: License Compatibility Check

For each file in the skill's references/ and scripts/ directories:

1. Check for license headers or declarations within the file
2. Check for LICENSE files in the same directory
3. If the file contains content from a known standard (reporting guidelines, clinical scores, etc.), identify the source and its license

Classify each file with ${CLAUDE_SKILL_DIR}/references/license-compatibility-matrix.md (written for an MIT target; it also lists common checklist and package licenses). Bundle compatible content with the attribution header or license notice it requires; convert non-compatible content to the matrix's URL Reference Pattern; mark GPL/LGPL tools as optional external dependencies; treat an unknown license as incompatible (remove, or get permission).

Output

Present license audit table:

| File | Source | License | Compatible? | Action |
|------|--------|---------|------------|--------|

---

Phase 5: Validate and Test

Structural Validation

1. YAML frontmatter: Parse and verify the required fields (name and description per the Agent Skills spec, plus any field the target package's own validator requires)
2. File references: Every ${CLAUDE_SKILL_DIR}/... path resolves to an actual file
3. Script executability: Scripts in scripts/ have appropriate shebangs
4. Line count: SKILL.md should be under 500 lines for optimal loading
5. Description quality: Description should start with a verb and include trigger keywords

Final PII Re-check

Run audit_skill.sh --strict one final time on the cleaned skill. Must return exit code 0.

Cross-Platform Adapter Review

Check whether the skill can run in common desktop-agent environments:

Platform	Check
Claude Code	No hardcoded dependency on private ~/.claude paths unless documented.
Codex	SKILL.md is self-contained and installable under ~/.agents/skills/.
Cursor	A short .cursor/rules/*.mdc adapter can point to the canonical SKILL.md.
Windows	Commands avoid Unix-only assumptions or provide PowerShell/Python alternatives.
macOS/Linux	Shell examples use portable paths where possible.

If the package is intended for a workshop or classroom, read ${CLAUDE_SKILL_DIR}/references/classroom-distribution.md and follow it (direct-download ZIPs, release assets, classroom checklist).

README Entry Draft

Generate a table row matching the target package's README format:

| **{skill-name}** | {One-sentence description of what the skill does.} |

User Testing

Instruct the user to:

1. Copy the cleaned skill to a test location: cp -r <cleaned_skill> ~/.claude/skills/<skill-name>
2. Restart Claude Code
3. Test the skill triggers by typing /<skill-name> or relevant trigger phrases
4. Verify all phases work end-to-end on a sample input

Gate: User confirms testing is complete.

---

Phase 6: Package and Commit

Copy to Target Package

cp -r <cleaned_skill_path> <target_package>/skills/<skill-name>/

Update README

Apply the README entry drafted in Phase 5:

• Add row to the appropriate table (Available Now / Coming Soon)
• Update pipeline diagram if the skill adds a new stage
• Update skill count if mentioned in prose

Generate Commit Commands

Present the exact commands but do NOT auto-execute push:

cd <target_package>
git add skills/<skill-name>/
git add README.md
git diff --cached   # User reviews
git commit -m "Add <skill-name>: <one-line description>"

Gate: User reviews git diff --cached and explicitly approves the commit. Push is always manual.

Post-Publish

Remind the user to:

• Add the skill to any marketplace listings if applicable
• Test installation from a clean clone: git clone <repo> && cp -r <repo>/skills/<skill-name> ~/.claude/skills/
• For classroom distribution, finish the release steps in references/classroom-distribution.md.