Skill Creator King

The Specific Problem It Solves

When building AI Agents, managing Skills—which are Markdown files defining triggers, workflows, and instructions—is often a fragmented process. Development lacks unified standards: starting from scratch involves no scaffolding, quality reviews rely on subjective judgment, manual fixes risk introducing new defects, and version iterations lack traceability. This leads to several pain points:
- Inconsistent Quality: Without systematic auditing, common anti-patterns (like YAML syntax errors, version inconsistencies, missing files) recur.
- High Fix Risk: Direct manual edits can break structure, and rolling back to a safe state is difficult.
- Disconnected Workflow: Tools for design, build, and verification are not cohesive, leading to information drift across files.

Skill Creator King (SCK) is a static analysis toolchain and workflow assistant designed specifically for these issues. It does not replace LLMs or human judgment for functional logic; instead, it focuses on the formal correctness of Skill files and process standardization through an automated toolset, structuring the entire lifecycle management.

How It Works: Core Capabilities & Key Steps

SCK’s core positioning is a Linter that "checks form, not function." It provides a complete five-stage workflow centered around four core capabilities: creation from dialog, verification via scoring, automated fixing, and iterative evolution.

Core Capabilities & Toolchain

• Template Scaffolding (init_skill.py): Generates a standardized directory structure, ensuring new Skills include all required files (e.g., SKILL.md, README.md, CHANGELOG.md).
• Structured Validation & Scoring:
• validate.py: Performs basic checks for file completeness, version consistency, and scans for anti-patterns (defined in anti-patterns.yaml), outputting a structured list of issues.
• quality-audit.py: Conducts a 14-dimension quality audit, assigning a grade (A/B/C/D). Dimensions cover documentation structure, version sync, error handling, etc., with configurable dimension skipping for different platforms (e.g., WorkBuddy, OpenClaw).
• Safe Fixing Framework (autofix/engine.py): A key design. Fixes are not ad-hoc but follow a "backup → LLM fix → verify" safety skeleton. A snapshot backup is mandatory before any modification. After fixing, validation is re-run to ensure the score doesn’t drop; if it does, --rollback restores the backup. This decouples auditing (read-only) from fixing (read/write), communicating via structured JSON.
• Semantic Consistency Check: During Phase 4.1, SCK instructs the LLM to cross-compare SKILL.md, DESIGN.md, README.md, and other files for consistency in stage counts, numbers, deliverable lists, and feature descriptions, preventing semantic drift.

Key Workflow Steps

The process strictly follows Phase 0 (Pre-flight) to Phase 5 (Evolve), with the verification stage (Phase 4) being particularly detailed:
1. Structure Scan: Run validate.py --json to get a base list of issues.
2. Quality Baseline: Run quality-audit.py for an initial score.
3. Safe Fix: Run engine.py --backup to create a snapshot, then the LLM fixes files sequentially based on the validate.py output.
4. Fix Verification: Run engine.py --verify, which re-executes validate.py and quality-audit.py to confirm all issues are resolved and the score hasn’t dropped. If it drops, --rollback is triggered.
5. Deep Review (Optional): For major version changes, enable quality-audit.py --llm-review, where the LLM answers four fixed questions on semantic drift, terminology consistency, etc.
6. Self-check & Testing: After any modification, automated scans are mandatory. Ensure new files are referenced by workflows, scripts handle success/rate-limit/failure paths, and auto-generated tests/test.py is executed.

Applicability & Important Caveats

SCK is a powerful static quality control tool, but its boundaries must be clear:
- It’s a Linter, Not a Runtime: It only checks Skill files for format, structure, and documentation consistency. The correctness of functional logic, workflow reasonableness, and trigger effectiveness must be validated via actual LLM execution or manual testing.
- Platform Variability: Different AI Agent platforms (e.g., WorkBuddy vs. OpenClaw) have different maximum scores and dimension checks. The target platform must be confirmed in Phase 0, and quality-audit.py adapts automatically.
- LLM Dependency for Fixes: The fixing phase relies on the LLM’s direct file read/write capabilities. Safety nets (backup/rollback) exist, but fix quality depends on the LLM’s comprehension.
- Strict Process Adherence is Mandatory: After any modification to Skill files, validate.py and quality-audit.py --no-cache must be re-run for a full scan to ensure a zero-issue state, maintaining version consistency discipline.