Version: 4.0 (Ecosystem-Aware Design, Cross-Skill Delegation, Mermaid Decision Trees)
SKILL.md file.
- Google Antigravity: Native Support. (Perfect match for YAML-Frontmatter & Progressive Disclosure).
- Cursor IDE: Native Support. (Agents correctly read XML tags and execute the 4-Phase routing).
- Claude Code / Root CLIs: Native Support. (The Dedup-Scan is restricted to local queries to prevent runaway
find /commands).
The Skill-Engineer is a meta-agent. It is not designed to assist end-users directly with domain-specific tasks. Instead, it develops, audits, and configures other agents (skills). It forces the underlying LLM into methodological rigor, eliminates generic fluff, and prevents "Yes-Man" behavior.
Credits & Architecture: The implementation of the "Progressive Disclosure Architecture" (offloading context into
references/andscripts/to conserve the Context Window) is heavily based on the concepts of the Skill-Creator by Clawdbot (Developer: Peter Steinberger).
The Skill-Engineer is a Stateful Agent. It maintains a <configuration> block directly within its own codebase.
- On Initial Launch (
STATUS: UNCONFIGURED): The agent blocks all other workflows. It detects its current runtime environment (e.g., Platform = Antigravity, LLM = Gemini, Language = English) and proposes these as system defaults. Once the user confirms, the agent uses its file-system permissions to physically rewrite the block in its ownSKILL.mdtoCONFIGURED. - In Normal Operation (
STATUS: CONFIGURED): The agent stops prompting the user for parameters (Target LLM, Platform, etc.) and silently enforces its hardcoded defaults as law. This saves tokens and dramatically reduces user friction. - Command
/config: The user can type/configat any time to explicitly overwrite the stored defaults in the file system.
Before the agent starts generating, it performs a two-step analysis:
- Stateful Skill-Map Check: Checks via
.skill-config.jsonwhether anECOSYSTEM.mdmap exists. If it is missing, the Engineer delegates to its specialized sub-agent (skill-mapper) to recursively map the workspace ecosystem. This architectural split (Separation of Concerns) offloads heavy scanning logic into a dedicated tool. - Pruning & Router-Check: The Engineer reads only the
ECOSYSTEM.mdmap (instead of blindly opening all 50 files in the folder), which saves tokens and prevents context inflation. If it detects direct conflicts through the map, it automatically suggests generating a "Router-Skill". - Benchmark Scan (Optional): For complex skills, the agent can research external solutions to identify must-have features and USPs.
If there is no duplication issue (or the user explicitly requests a rebuild), the intelligent routing takes over:
- Mode A (Blank Canvas): The user only has a rough idea. The agent initiates the full 10-question interview.
- Mode B (Existing Draft): The user provides a skill draft. The agent performs a hard Gap Analysis.
- Mode C (Audit / Edit): For existing skills.
- C-Audit: The agent evaluates a legacy skill against the 13-point Quality Gate and presents a report.
- C-Edit: The agent executes a defined detail change immediately and validates only the modified section.
- Mode D (Quick Draft): Trivial utility skills skip the interview — instant light template.
- Mode E (Regeneration): (New in v4.0) An extensive planning document, strategy paper, process document, SOP, or knowledge base exists, but no SKILL.md. The agent reads the legacy document, extracts core functions, workflows, rules and taboos, and translates them into the SKILL.md architecture. Key difference from Mode B: Mode B has a skill draft. Mode E has domain knowledge in unstructured form.
The Skill-Engineer is programmed not to be easily satisfied. Before yielding the final skill (strategy proposal) in Phase 3, it is strictly required to formulate a counter-argument (Veto) against its own approach. If no logical flaw exists, it must instead identify the most dangerous operational Edge Case where the future target skill is most likely to fail.
When the Interview identifies that sub-tasks could be handled by existing skills, the Engineer injects a Delegation Pattern into the generated skill:
- A
NEVER ... delegate to [skill-name]rule in<operational_rules>with a standardized briefing template. - An expected return format specification.
- A corresponding step in
<process_workflow>where the delegation occurs.
Example: The google-ads-manager skill delegates copywriting to ad-expert and brand checks to brand-guidelines.
Before the Skill-Engineer releases a target skill, it silently forces it through this checklist (13 maximum points across 4 categories):
- Native XML-Tags (no prose instructions)
- Mechanics over Adjectives
- Platform-agnostic format (YAML-Frontmatter)
- Environment Check in Step 1
- Self-Correction / Chain-of-Thought injected
- Explicit Taboo list
- Unconfirmed references marked with
<!-- UNCONFIRMED -->
- Zero-Shot Triggering precision
- Few-Shot example included
- Stateful/Stateless correctness
- Cross-Skill Delegation: Sub-tasks delegated to specialized neighbor skills with briefing + return format.
- Reference Quality: Every reference file independently usable, under 200 lines, Mermaid for decision trees with ≥3 branches.
- Update Capability: Dynamic reference files (inventories, lists) contain date, maintenance rules, and status tracking.
(This document is an automatically generated system description of the agent)