Second Brain Router

Purpose

This is a standalone lightweight trigger skill for a personal second brain.

It routes a user question through cognitive anchors, local search, and experience memory without loading a large wiki-maintenance skill by default. It can be shared as one skill because it includes the minimum writeback protocol itself.

Expected Vault Shape

Assume a Markdown wiki or vault with these optional paths:

• my-core-questions.md — long-term questions
• links/index.md — cognitive anchor index
• links/*.md — type: anchor cognitive anchor pages
• concepts/ — stable concepts
• queries/ — reusable answers and saved question explorations
• practices/ — action routines and workflows
• _meta/automation-runs/ — automation reports

For Dongge's machine, default paths are:

• Hbrain repo: /Users/jianghaidong/Hbrain
• Wiki root: /Users/jianghaidong/Hbrain/llm-wiki
• Anchor index: /Users/jianghaidong/Hbrain/llm-wiki/links/index.md

If these paths do not exist, infer the local wiki root from the current project or ask one concise question.

Trigger Classifier

Classify each user request:

1. ordinary: answer normally; do not run the second-brain loop.
2. anchor-worthy: run the lightweight second-brain loop.
3. writeback-needed: run the lightweight loop, then write back only if the user explicitly requested saving/updating.

Use anchor-worthy when the user asks about:

• understanding, judgment, strategy, reflection, decision, writing, output, or review
• AI agents, coding agents, second brain, personal knowledge systems, memory, CASS, Gbrain, wiki
• cognition, metacognition, long-termism, compounding, feedback loops, self, no-self, meaning
• any long-term core question in the user's vault

Stay ordinary for simple commands, transient facts, basic calculations, or code tasks unrelated to the knowledge system.

Use writeback-needed only when the user says or clearly implies: save, write to wiki, update anchor, create page, preserve this, 沉淀, 保存, 写入 wiki, 更新锚点, 创建页面, 整理到第二大脑.

Lightweight Loop

For anchor-worthy and writeback-needed requests:

1. Read only lightweight routing files when they exist:
   • my-core-questions.md
   • links/index.md
2. Select 1-3 likely anchors from links/*.md.
3. Read only those selected anchor pages.
4. Run focused recall when available:
   • gbrain search "\x3Cuser question>" --limit 5
   • gbrain search "\x3Canchor name>" --limit 5
   • cm context "\x3Cuser question>" --json --limit 5 --history 5
5. If a recall tool fails or is unavailable, continue and state the failure briefly.
6. Answer using:
   • the user's first-brain framing
   • the anchor minimal model
   • retrieved local context
   • CASS/experience lessons when available
7. Decide whether to write back or propose a writeback.

Do not read the whole wiki index, full log, full schema, large references, or raw sources during the lightweight loop.

Minimal Anchor Template

When explicitly creating or normalizing an anchor, use:

---
title: Anchor Name
created: YYYY-MM-DD
updated: YYYY-MM-DD
type: anchor
tags: [anchor, contact-point]
sources: []
weight: 0
last_active: null
core_questions: [Q1]
status: active
---

# Anchor Name

## 一句话理解

## 我的大脑里必须记住的最小模型

## 什么时候调用它

## 连接到第二大脑

## 常用提问

## 最近调用记录

## 下一步要补的连接

Writeback Rules

Default to no file edits.

If the user explicitly asks for writeback, first choose the target layer:

• links/ for cognitive anchors
• concepts/ for stable ideas
• queries/ for reusable answers
• practices/ for action routines
• _meta/automation-runs/ for automation reports
• CASS/playbook only for repeated agent workflow lessons

Keep writeback small and auditable:

1. Read the target file if it exists.
2. Preserve existing content and frontmatter.
3. Add only the smallest durable update.
4. Bump updated dates when editing wiki pages.
5. Add or propose an index/log update when the local wiki convention requires it.

Never delete, rename, migrate, publish, send external messages, edit raw source files, or handle sensitive raw data without explicit user confirmation.

Routing Receipt

For every anchor-worthy or writeback-needed answer, include a concise receipt:

第二大脑回路：
- 模式：anchor-worthy / writeback-needed
- 长期问题：Q?
- 锚点：[[links/...]], [[links/...]]
- 召回：Gbrain / CASS / anchor-only / unavailable
- 写回：不需要 / 建议 / 已按要求写入

If the request is ordinary, do not include the receipt unless the user asks whether the loop was used.

Anti-Patterns

• Do not load a heavy wiki-maintenance skill by default.
• Do not do full-vault scans for ordinary questions.
• Do not create wiki pages just because a question is interesting.
• Do not hide recall failures; short transparency is better than pretending the loop ran.
• Do not answer as generic internet knowledge when a relevant anchor exists.