AgentSkillsCN

skill

技能

SKILL.md

Personal Knowledge System

Overview

This skill provides access to my personal knowledge base - a distilled collection of insights, decisions, and learnings from past conversations. Use it to maintain consistency across conversations and avoid re-explaining past decisions.

Available Tools

1. get_index

Returns a compressed overview of all topics and projects.

When to use:

  • •At the start of technical conversations
  • •When asked "what do I know about X?"
  • •When asked "what are my active projects?"

Example triggers:

  • •"What are my current projects?"
  • •"What topics have we discussed?"
  • •"Show me my knowledge index"

2. get_context(topic)

Returns the current view and key insights for a specific topic.

When to use:

  • •When discussing a specific technical topic
  • •When I mention a project by name
  • •When you need my established position on something

Example triggers:

  • •"What's my view on MLX?"
  • •"Tell me about the trading system project"
  • •"What do I think about React state management?"

3. get_deep(id)

Returns the full entry with all evidence and evolution history.

When to use:

  • •When asked about how my thinking evolved
  • •When provenance/sources matter
  • •When there's a contested entry needing resolution

Example triggers:

  • •"How did my view on X evolve?"
  • •"Walk me through the decisions on project Y"
  • •"What evidence supports this view?"

4. search(query)

Semantic search across all knowledge.

When to use:

  • •When topic/project name is unclear
  • •When looking for related discussions
  • •When asked "have we discussed X?"

Example triggers:

  • •"Have we talked about volatility trading?"
  • •"Find anything related to model fine-tuning"
  • •"What do I know about Python packaging?"

Routing Logic

code
User message mentions specific topic/project
  → get_context(topic) first, get_deep(id) if details needed

User asks about general knowledge state
  → get_index()

User asks about evolution or provenance
  → get_context(topic) then get_deep(id)

User asks if we've discussed something
  → search(query)

Entry is contested (shown in index)
  → get_deep(id) to show both positions

Important Behaviors

  1. •Contested entries have multiple positions - present BOTH views, don't pick one
  2. •Compressed entries have has_full_content: false - mention that archive exists
  3. •Evolution history shows how thinking changed - useful for explaining past shifts
  4. •Evidence links back to specific conversations - use for credibility

Response Format

When referencing knowledge:

  • •Cite the entry domain/name naturally
  • •Mention confidence level if relevant
  • •Flag if entry is contested or stale
  • •Offer to get_deep if user wants provenance

Example Session

User: What's my current view on MLX layer selection?

Assistant: calls get_context("MLX layer selection")

Based on your knowledge base, your current view on MLX layer selection is:

"For LoRA fine-tuning on MLX, select layers based on task specificity - earlier layers for domain adaptation, later layers for output format changes. The default all-layers approach often over-fits."

Key insights:

  • •Layer selection significantly impacts training efficiency
  • •You've had success with selective layer training on code models

This is a high confidence active entry. Want me to pull the full evolution history?