Skill Sync

Skill Sync

Detect and fix drift between a package's code and the skills that describe it.

Why this matters

Skills are documentation for AI agents. Like any documentation, they rot when the code they describe changes. A stale skill is worse than no skill — it confidently tells the agent to do things that no longer work.

When to sync

• After significant code changes (new features, renamed functions, changed

signatures, removed modules)

• Before a release (skills should match the version being shipped)
• When a user reports that a skill's instructions don't match reality
• Periodically as maintenance (e.g., monthly, or when touching the package)

Process

1. Inventory current skills

Find all skills associated with this package:

# Check project-local skills
ls -la .claude/skills/

# Check if the package ships skills in data/
ls -la {pkg_name}/data/skills/ 2>/dev/null

# Check for symlinks from the global store
skill list-skills

2. Diff skills against code

For each skill, read its SKILL.md and check every claim against the current codebase:

API references:

• Do the functions/classes mentioned still exist? (grep for them)
• Have their signatures changed? (read the source)
• Have default values or behaviors changed?
• Are there new parameters the skill should mention?

Code examples:

• Do the code examples still run correctly?
• Do imports still resolve?
• Are there deprecated patterns the skill still recommends?

File paths:

• Do referenced file paths still exist?
• Have modules been moved or renamed?

Behavioral claims:

• Does the skill describe behavior that has changed?
• Are there new edge cases or gotchas?

3. Categorize findings

Group issues by severity:

• Breaking — Skill references something that no longer exists or has changed

in a way that would cause errors (wrong function name, removed parameter, changed return type). Fix these immediately.

• Stale — Skill is technically correct but incomplete (new features not

mentioned, new options not documented). Update when convenient.

• Cosmetic — Minor wording issues, formatting. Low priority.

4. Apply updates

Edit each skill's SKILL.md to fix the issues found. When updating:

• Keep the existing structure and style — don't rewrite the whole skill.
• Update specific claims, examples, and references.
• Add notes about new features or changed behavior.
• Remove references to things that no longer exist.
• Update the description in frontmatter if the skill's scope has changed.

After editing, validate:

skill validate .claude/skills/{name}

5. Report

Summarize what changed:

## Skill Sync Report

### my-skill
- FIXED: `search()` now takes `backends` parameter (was `sources`)
- ADDED: New `link_skills()` function documentation  
- REMOVED: Reference to deprecated `AGENT_TARGETS` dict (now a Registry)

### other-skill  
- No changes needed ✓

Automation tip

For packages with CI, consider adding a step that checks skill freshness:

from skill import validate
import subprocess

# Validate all skills
for skill_dir in Path('.claude/skills').iterdir():
    if skill_dir.is_dir():
        issues = validate(str(skill_dir))
        if issues:
            print(f"⚠ {skill_dir.name}: {issues}")

This won't catch semantic drift (the skill says "returns a list" but now it returns a generator), but it catches structural issues like missing fields.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: thorwhalen
• Source: thorwhalen/skill
• License: MIT
• Homepage: https://thorwhalen.github.io/skill/

Install and usage instructions live in the source repository linked above.