Structured Agent Skill Framework for LLM Collaboration: Knowledge Base and Task Decomposition Patterns

This article introduces bukzor-agent-skills, a collection of Agent skills for LLM-human collaboration. It primarily provides structured patterns for knowledge base management, document workflows, and task decomposition, supporting multi-Agent collaborative work. The project adopts an "Opinionated" design philosophy, focusing on best practices in specific scenarios to avoid the ambiguity and inconsistency of one-size-fits-all tools.

With the deep application of LLMs in software development, knowledge management, and other fields, building structured collaboration patterns has become a key issue. Traditional tools often lead to ambiguity and inconsistency due to their pursuit of generality. Through its "Opinionated" design, this project provides clear usage boundaries and guiding principles for specific scenarios, addressing the lack of structure in LLM-human collaboration.

llm-kb: Structured Knowledge Base Pattern

• Core principles: Small file focus (single topic), separate maintenance guidelines (CLAUDE.md), on-demand summarization
• Directory structure: Centered on the .kb/ directory, including root-level maintenance guidelines, category-level summaries, etc.
• Upgrade signals: Plural filenames, parallel chapters of the same type, list-based main structure, independent lifecycle management

Other modules: llm-collab (collaboration workflow specification), llm-subtask (task decomposition), llm-design-kb (design decision knowledge base), llm-must-read-kb (priority content)

• Pre-dependency declaration: Use requires: and depends: directives to clearly specify files to read before taking action, avoiding context loss
• Pattern validation: Supports JSON Schema validation for YAML metadata files to prevent format drift
• Naming conventions: kebab-case lowercase naming, descriptive filenames, zero-padded number prefixes (e.g., 001-setup.md)

• Multi-Agent collaborative development: The unified .kb/ structure allows Agents to quickly locate domain knowledge
• Long-term project maintenance: Separation of maintenance guidelines and content reduces document update costs
• Knowledge transfer: New Agents quickly establish context through the structured knowledge base, reducing reliance on verbal communication

Core insight of the project: LLM collaboration tools should respect the cognitive boundaries of humans and AI, extending memory through structured information rather than increasing burden. Key principles:

1. Explicit over implicit (explicit declaration of dependencies and rules)
2. Separation of concerns (separation of maintenance guidelines and content)
3. Progressive complexity (from simple files to directory structures)
4. Machine readability (naming and structure adapted for LLM parsing)

bukzor-agent-skills provides a practice-tested pattern language for LLM-human collaboration, with core value lying in its structured, explicit, and maintainable design philosophy. As AI-assisted development becomes normalized, such frameworks will play an important role in software engineering. It is recommended that teams wishing to establish scalable AI collaboration processes conduct in-depth research on this project.