NOTE

Agent Skills Index

authorclaude-sonnet-4-6 aliasesskill-directory, agent-capabilities titleAgent Skills Index statusactive date2026-05-02 typepermanent

Agent Skills Index

A directory of specialized skills and procedural knowledge modules used by agents (Gemini, Claude, etc.) to maintain the vault and execute workflows.

Core Skills

  • Session Distillation: Triggers on "distill" or "capture session". Responsible for converting temporary conversation context into durable Permanent Notes.
  • Vault Maintenance: Procedural knowledge for adhering to YANP and managing note lifecycles.

.NET Agentic Stack (Tier-1, Planned)

> These five notes are planned but do not yet exist in the vault. Links are aspirational stubs pending synthesis from source material in 00_Raw/.

  • dotnet-agent-integration — Main architectural bridge for .NET/C# in agentic loops.
  • csharp-mcp-sdk — Building type-safe, performance-critical MCP servers.
  • lm-kit-dotnet — On-device inference and GGUF integration for .NET agents.
  • foundry-local — Local ONNX inference with OpenAI-compatible interfaces.
  • microsoft-data-sqlite-agent-patterns — Lightweight, idempotent memory services.

Implementation Pattern

Skills are injected into an agent's context at session start or loaded on demand. Three concrete forms in this fleet:

  • Slash-command skills (Claude Code): A skill file (e.g., session-distillation.md) loaded by the harness when the user types the corresponding slash command (e.g., /distill). The skill text becomes part of the active context and instructs Claude to call save_memory() with a structured summary.
  • Instruction-set skills (Gemini CLI): A named section in the system prompt or GEMINI.md referencing vault conventions — e.g., the YANP compliance block that instructs Gemini to call write_file() with correct frontmatter whenever it synthesizes a new note.
  • PowerShell scripts (Codex): Automation scripts in 02_System/ (e.g., audit-yanp.ps1) that encode vault health rules as executable checks, run by the Codex engineer on demand or in CI.

All three forms encode the same conventions; they differ only in the native tool API each model exposes.

Skills vs. Knowledge

There is a critical distinction between active agent Skills and passive Vault Knowledge:

Feature Native Skill (Active) Vault Note (Passive)
Role Active Rules: Instructions injected into the agent's core system prompt. Passive Facts: Data stored in the vault for the agent to reference.
Activation System-prompt injection: Prepended to the agent's context at session start, or loaded on demand via a slash command (e.g., /distill loads the distillation skill, which then calls save_memory()). Manual: The agent must find it via search or be directed to it.
Capabilities Tool-Aware: Written to use specific agent tools (e.g., save_memory, write_file). Convention-aware prose: Human-readable text that requires an agent configured for this vault's YANP conventions to use correctly; not universally parseable by arbitrary models.

Why We Need Both

> Status: Active — This pattern is in current use across all three fleet members (Gemini, Claude, Codex).

To maintain a consistent vault across different models, we use a Native Wrapper strategy:

  1. Vault Notes define the conventions (the "Instruction Manual").
  2. Native Skills implement those conventions using model-specific tools (the "Executable Code").

The strategy exists because each model exposes different tool APIs (save_memory vs. write_file vs. PowerShell) — a single YANP-compliant outcome requires model-specific wiring around a shared convention.

References