Agent Skills Architecture

Agent Skills Architecture Standard

Priority: P0 (CRITICAL)

The primary goal is Maximum Information Density and Automated Precision.

🏗️ Core Architectural Pillars

1. Separation by Package (Granularity)

• Rule: Separate skills based on specific library/framework dependencies.
• Goal: Avoid context pollution. Don't load "Riverpod" instructions into a "BLoC" project.
• Example: flutter/bloc-state-management vs flutter/riverpod-state-management.

2. Presence = Active (Simplified Configuration)

• Logic: If a skill is listed in .skillsrc, it is considered active.
• Standard: Remove legacy enabled flags. Control activation via inclusion/exclusion lists.

3. CLI Detection & Dynamic Exclusion

• Mechanism: The CLI (ags) maps package.json/pubspec.yaml dependencies to skill IDs.
• Exclusion: Irrelevant sub-skills are automatically added to the exclude list during initialization if their corresponding packages are missing.

4. Progressive Disclosure (Three-Level Loading)

• Level 1 (Metadata): Triggers activation via high-precision keywords/files.
• Level 2 (SKILL.md Body): Core imperative logic (<100 lines). No conversational fluff.
• Level 3 (References): Detailed examples, complex patterns, scripts. Lazy-loaded on-demand.

📦 ID & Naming Standards

• Category: Lowercase letters (e.g., flutter, nestjs).
• Skill ID: Kebab-case, must match the directory name.
• Registry ID: Must match the Skill ID for automated CLI detection/exclusion.

📋 High-Density Writing Style

• Imperative Mood: Use "Use X", "Avoid Y". No "Please" or "You should".
• Token Compression: Skip articles ("the", "a") where possible. Use bullet points > paragraphs.
• Anti-Patterns: Avoid conversational intros ("In this skill, we see...").

🔗 Internal References

• Skill Creator Standard - Detailed authoring rules.
• Resource Organization - Folder structure best practices.