Wiki Index

Content catalog. Every wiki page listed under its type with a one-line summary. Read this first to find relevant pages for any query. Last updated: 2026-07-17 | Total pages: 29

Entities

• gbrain — Personal knowledge brain for AI agents by Garry Tan. Two-repo architecture, synthesis layer, self-wiring knowledge graph.
• claritys-system-prompts — Transparency project collecting system prompts from major AI models. Key reference: Claude Design prompt.
• oh-my-hermes — Agent team architecture for Hermes. 23 skills, 6 specialized agents, kanban coordination, autonomous CTO loop.
• personal-os — Goal-driven task management for AI. Brain dump inbox, priority system (P0 max 3, P1 max 7), session evals.
• hindsight-agent-memory — Vectorize.io 嘅 agent memory system. 仿生結構（World/Experiences/Mental Models），4 策略並行檢索，LongMemEval SOTA。需要 LLM API key。

Concepts

• ai-marketing-workflow — 六步 AI Marketing Workflow：Research → Strategy → Creative Brief → Activation → Conversion → Optimization。AI = Acceleration Layer，Human = Decision Gates。
• loop-skill — Filesystem-backed validator-gated loop for auditable, trackable AI batch outputs.
• agent-knowledge-architecture — Brain vs Agent separation pattern. Execution layer separate from knowledge layer. Strategic moat = accumulated knowledge.
• claude-design-workflow-patterns — Reusable design agent workflow. Context-first, variation generation, tweak model, progressive complexity.
• agent-team-pattern — Specialized multi-agent coordination. Orchestrator + specialists + shared state + human at decision points only.
• hermes-profiles-multi-user — Profile system vs personality overlay for multi-user/multi-persona setups. True isolation vs tone switching.
• preschool-development-photo-voice — 3–5 year old developmental domains mapped to photo+voice learning. Reusable knowledge for child-related product design.
• multi-agent-hierarchy-brakes — Multi-agent Butler/Assistant hierarchy with database-level safety brakes (depth, concurrency, round limits). Atomic SQL enforcement, not prompt-level rules.
• structured-memory-cards — Three card types (State/Event/Knowledge) for cross-session AI memory. Cardinality, half-life decay, conflict arbitration, time anchoring.
• llm-randomness-as-feature — LLM stochasticity is a feature, not a bug. Design deterministic fences (code/DB), not soft prompt guardrails. The asymmetry argument: fence cost is low, protection is permanent.
• llm-compliance-audit — Three-layer compliance audit for LLM tool use: Skill Contract, Runtime Probe, Post-Hoc Audit. Five failure modes (Skill Drift), Stratified RAG for tool selection, three-stage gentle intervention.
• self-evolution-safety — Can AI modify its own source code safely? Four-layer model, three lines of defense, sandbox architecture, blue-green deployment. Kill-switch OFF by default.
• loop-engineering-hooks — Loop Engineering Hook 事件驅動自動化：Prompt → Skill → Hook → Loop。6個實用 Hook 玩法，從對話到循環。
• hermes-hook-system — Hermes Hook System 完整參考文檔：三套系統（Gateway Events, Shell Hooks, Webhook Subscriptions），可用事件、配置方法、實用範例。
• harness-engineering-claude-code-distillation — Claude Code 源碼蒸餾實踐：文件系統協調層、品味注入（基向量投影）、多 Agent review-action 循環、severity 收斂追蹤。
• ai-era-must-read-books — AI時代必讀10本書：湧現、反饋、系統思維、數據思考、媒介理論、反脆弱、領導力、故事力、哲學懷疑、戰略思維。含閱讀渠道。
• sag-sql-retrieval-augmented-generation — SAG：SQL JOIN 動態構建事件關係，取代 RAG 向量模糊匹配同 GraphRAG 預建圖譜。快 + 深 + 增量更新。Recall@K 新 SOTA。
• nextjs-proxy-pitfall — Next.js SPA 唔适合經 Flask proxy（60s+ latency）。每個 JS chunk 都要過一層 proxy，加埋 latency 就爆炸。
• zeabur-networking — Zeabur 每個 port 需要單獨 domain generation。Private ports 只供 internal，Public 先 external。
• pm2-esm-pitfall — 當 package.json 有 "type": "module" 時，PM2 ecosystem config 需要 .cjs extension。
• container-pid1-pitfall — Container PID 1 唔可以 kill — 會連帶殺死所有 child processes（包括 PM2）。
• open-design-tools-dev — Open Design 嘅 lifecycle control plane 係 tools-dev，唔係 npm scripts。路徑: ./tools/dev/bin/tools-dev.mjs。
• server-resource-management — VPS RAM budgeting lessons. Next.js dev server = 2.5GB+, daemon-only = ~475MB. Monitor, budget, prefer lightweight modes.

Comparisons

• open-source-tts-2026 — VoxCPM2 vs Qwen3-TTS vs MiMo-V2.5-TTS vs CosyVoice3: benchmark, features, language support, deployment options.

Queries