Open Agent Skills Standard Compliant

The Production Framework for
Self-Improving Knowledge Vaults

AIのドリフトとハルシネーションを止め、カオスなAIメモを機械的に検証された自己組織化セカンドブレインへ。頻出する修正は自動的に再利用可能なスキルへ昇格。

bash — npx llm-wiki-loop init

🧬 自己組織化ライフサイクル

Rotate through the 4-stage closed loop that powers autonomous vault evolution.

01

1. Immutable Ingestion (Raw)

Never edited by LLM. Original notes, benchmark runs, papers, and logs are stored as ground truth.

02

2. Grounded Compilation

AI agent compiles raw sources into concise wiki pages with mandatory Raw: provenance links.

03

3. Mechanical Invariant Check

check_evidence.py verifies every number, date, and quote against raw bodies. 0-error rule.

04

4. Self-Organizing GC & Auto-Skillify

Outdated facts are archived (never deleted). Repeated fixes (2+) automatically evolve into reusable skills.

STAGE 1 : IMMUTABILITY

Source of Truth Ingestion

All human thoughts, logs, and external data land in raw/notes/ and remain immutable forever.

# raw/notes/2026-08-16-benchmark.md > Source: Performance Team Published on 2026-08-16. Latency: 15.0ms under 1,000 requests.
STAGE 2 : GROUNDED COMPILATION

Provenance Linking

Agents synthesize insights and attach explicit provenance links. No claim without evidence.

# wiki/topics/latency-profile.md > Raw: [Benchmark](../../raw/notes/2026-08-16-benchmark.md) Latency is 15.0ms at 1,000 req scale.
STAGE 3 : MACHINE LINTER

0-Hallucination Evidence Check

Strict mechanical matching checks candidate dates, grouped numbers, and quoted strings.

❯ npx llm-wiki-loop check # Evidence check 0 fidelity suspects, 0 evidence errors ✅ 100% Mechanically Grounded
STAGE 4 : AUTO-EVOLUTION

Autonomous Skill Promotion

When a fix appears twice in log.md, the system proposes promoting it into a permanent skill.

Auto-Skillify Trigger: Pattern detected 2+ times in log.md 👉 Proposed: .agents/skills/fix-db-timeout/

🌐 1-Click Multi-Agent Compatibility

Inject standardized skills into all your favorite AI coding assistants with zero configuration.

🤖

Claude Code

Direct integration with ~/.claude/skills for terminal agent pairing.

Auto-Detected

Cursor & Windsurf

Project-level .cursor/skills/ and rules support for seamless in-editor context.

Auto-Detected
🧠

Gemini & OpenCode

Full support for .gemini/skills and open agent specifications.

Auto-Detected

🚀 10秒で開始

Initialize your first self-improving knowledge vault right now.

npx llm-wiki-loop init
⭐ Star & Read Full Docs on GitHub