Second Brain Growth#

Overview#

Use this skill to evaluate the user's second-brain growth as a closed loop: structured deposition, connection, 人脑 recall, transformation into action/output, and debt control. Default to Chinese for user-facing summaries.

Core doctrine:

• Growth is not storage. Effective growth means new material becomes connected, recallable, and useful for judgment/action/output.
• Keep knowledge deposition and anchor training separate. Do not create anchor-events, modify anchor:, or infer 人脑 training just because a page was written.
• Treat anchor-events as evidence of human-side recall demand, not as a log of AI reading material.
• Use absolute dates in reports.

Roots#

• Hbrain repo: /Users/jianghaidong/hbrain
• Wiki root: /Users/jianghaidong/hbrain/llm-wiki
• Anchor dashboard: /Users/jianghaidong/hbrain/llm-wiki/_meta/anchors/index.md
• Core questions: /Users/jianghaidong/hbrain/llm-wiki/_meta/my-core-questions.md
• Anchor events: /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-events/YYYY-MM.jsonl
• Router misses: /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-events/router-misses/YYYY-MM.jsonl

Ecosystem Roles#

Use the same metric model across Codex, Hermes, and OpenClaw, but adapt the delivery contract:

RuntimeRoleOutput expectationCodexinteractive evaluator and skill maintainerconcise diagnosis in chat, optional file edits when askedHermesdaily/weekly scheduled operatordeterministic commands plus Markdown report; no prompt-side derived state editsOpenClawmulti-agent orchestration layertask handoff block, machine-readable summary, clear next-agent actions

Hermes is the daily operation interface. OpenClaw is the experiment/orchestration sandbox. Codex remains the engineering executor for changing this skill, scripts, or wiki automation.

Runtime Modes#

Codex Interactive#

• Use when the user asks a direct question in chat.
• Read local Hbrain sources, compute the five metric groups, and answer in Chinese.
• Avoid persistent writes unless the user asks for a saved report.

Hermes Cron#

• Use when the user asks to schedule, automate, or integrate this check into Hermes.
• Prefer hermes cron create --skill second-brain-growth --workdir /Users/jianghaidong/hbrain ....
• Hermes cron should orchestrate commands, read outputs, and write/append reports only. It must not hand-edit weight, last_active, hot, review-state, or anchor lifecycle fields.
• For persistent reports, write under:
  • Daily/weekly automation: /Users/jianghaidong/hbrain/llm-wiki/_meta/automation-runs/
  • Growth-specific audits: /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-governance/knowledge-growth/
• Include a stable final section named ## Agent Handoff so future Hermes/OpenClaw runs can pick up unresolved work.

Hermes cron prompt pattern:

bash

hermes cron create "0 21 * * 0" \
  --name "第二大脑有效增长周报" \
  --skill second-brain-growth \
  --workdir /Users/jianghaidong/hbrain \
  "Use $second-brain-growth to evaluate Hbrain effective knowledge growth for the last 7 days. Save a Markdown report if safe, and do not modify anchor lifecycle fields."

OpenClaw Orchestration#

• Use when the task should be split across main/design/code/coder/test or when OpenClaw is explicitly requested.
• The main/design agent should define the window, risk level, and output path.
• The code/coder agent may add scripts or run deterministic checks.
• The test agent should validate report shape, skill validity, and safety gates.
• Prefer OpenClaw --json output for automation, and include this skill name in the message body:

bash

openclaw agent --agent main --json --message "Use $second-brain-growth to audit Hbrain effective knowledge growth for 2026-06-08..2026-06-14. Return the Markdown report plus Agent Handoff. Do not edit derived anchor fields."

If OpenClaw is only used as an orchestrator, do not assume it has permission to mutate Hbrain. Treat edits as proposals unless the initiating prompt explicitly allows low-risk writeback.

Workflow#

1. Establish context.
   • Read the anchor dashboard and core questions before broad recall.
   • If the user gives a specific question, run dry routing only:

bash

python3 /Users/jianghaidong/.agents/skills/hbrain-cognitive-loop/scripts/hbrain_loop.py route --query "<user question>"

• Do not rerun with --record unless the user explicitly asks to count a human-side call, nomination, or miss.

1. Choose the window.
   • Default to the last 7 days for "this week" or an unspecified growth check.
   • Use month-to-date for "this month".
   • State the exact date range in the report.
2. Collect five metric groups.

A. 沉淀速度

• Count total Markdown pages.
• Count pages created/updated in the window from frontmatter and, when useful, git history.
• Segment by concepts/, queries/, practices/, comparisons/, entities/, and _meta/.
• Favor structural pages over raw volume.

Useful commands:

bash

rg --files /Users/jianghaidong/hbrain/llm-wiki -g '*.md' | wc -l
rg -l '^created: YYYY-MM-' /Users/jianghaidong/hbrain/llm-wiki -g '*.md' | wc -l
rg -l '^updated: YYYY-MM-' /Users/jianghaidong/hbrain/llm-wiki -g '*.md' | wc -l
rg --files /Users/jianghaidong/hbrain/llm-wiki -g '*.md' | sed 's#^/Users/jianghaidong/hbrain/llm-wiki/##' | awk -F/ '{print $1}' | sort | uniq -c | sort -nr

B. 连接速度

• Count total wiki links.
• Sample newly created structural pages and check whether each has at least two meaningful links to existing pages.
• Check whether new material is mapped to Q1-Q5 through direct links, anchor context, or weekly推进记录.

Useful command:

bash

rg -o '\[\[[^]]+\]\]' /Users/jianghaidong/hbrain/llm-wiki -g '*.md' | wc -l

C. 召回速度

• Parse anchor-events for used and recall-check.
• Count active anchors touched in the window.
• Treat recall_full, recall_partial, and recall_diverged as 人脑 training evidence.
• Include router misses as evidence of unmet recall demand.

Useful commands:

bash

wc -l /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-events/YYYY-MM.jsonl
wc -l /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-events/router-misses/YYYY-MM.jsonl
sed -n '1,120p' /Users/jianghaidong/hbrain/llm-wiki/_meta/anchor-events/YYYY-MM.jsonl

D. 转化速度

• Look for pages, weekly records, output dashboards, practices, code/project notes, decisions, and action experiments created from recent material.
• Report concrete examples rather than vague "learned a lot" summaries.
• Give special weight to evidence that knowledge changed judgment or behavior.

E. 负债压力

• Identify new isolated or underlinked structural pages.
• Note repeated router misses around the same theme.
• Note recall divergence, duplicate-like pages, stale high-priority anchors, and unresolved links only when they affect recall/action.
• Do not treat high orphan counts in the whole monorepo as automatically bad; focus on priority concepts, queries, and canonical anchors.

1. Score the loop.
   • Use the formula:

text

本周有效增长 =
  新增结构化页
+ 新增关键连接
+ 被调用/召回的锚点
+ 转化成输出或行动的案例
- 新增孤岛/重复/不可召回材料

• Prefer diagnostic labels over false precision:
  • 沉淀过快，召回滞后
  • 连接健康，转化不足
  • 召回活跃，沉淀不足
  • 结构均衡，进入复利
  • 负债上升，需要治理

1. Report in this shape:

md

**结论**
一句话判断增长状态和主要瓶颈。

**指标**
| 维度 | 本期证据 | 判断 |
|---|---:|---|
| 沉淀 | ... | ... |
| 连接 | ... | ... |
| 召回 | ... | ... |
| 转化 | ... | ... |
| 负债 | ... | ... |

**有效增长**
用公式解释本期是材料增长、网络增长、召回增长，还是行动增长。

**下周动作**
给 1-3 个最小可执行动作，优先修复瓶颈。

**Agent Handoff**
status: complete | needs-human | blocked
window: YYYY-MM-DD..YYYY-MM-DD
label: 沉淀过快，召回滞后 | 连接健康，转化不足 | 召回活跃，沉淀不足 | 结构均衡，进入复利 | 负债上升，需要治理
report_path: optional absolute path
safe_next_actions:
- ...
requires_confirmation:
- ...
blocked_by:
- ...

For Hermes/OpenClaw automation, also include this fenced machine-readable block at the end:

yaml

knowledge_growth_result:
  status: complete
  window: "YYYY-MM-DD..YYYY-MM-DD"
  label: "沉淀过快，召回滞后"
  report_path: null
  metrics:
    deposition: null
    connection: null
    recall: null
    transformation: null
    debt: null
  safe_next_actions: []
  requires_confirmation: []
  blocked_by: []

Interpretation Rules#

• If new/updated pages are high but anchor events and outputs are low, call it "沉淀速度快，召回/转化滞后".
• If new pages have weak links, prioritize connection repair before adding more material.
• If recall-check divergence rises, recommend compressing anchor capsules or morning recall, not adding more pages.
• If router misses cluster around one theme at least three times, propose a candidate anchor review; do not directly add anchor:.
• If output/action evidence appears, name the exact pages or projects and explain the loop it closes.
• If asked for an executive metric, use a 0-10 "有效增长指数" but always show the evidence behind it.

Compatibility Checks#

When upgrading or validating this skill:

• Run python3 /Users/jianghaidong/.codex/skills/.system/skill-creator/scripts/quick_validate.py <skill-dir>.
• Keep YAML frontmatter limited to name and description so Codex, Hermes, and OpenClaw skill readers can parse it.
• Keep agents/openai.yaml optional and UI-only; do not put required runtime logic there.
• Verify examples use current CLI shapes when changing ecosystem instructions:
  • hermes cron create --help
  • hermes --help
  • openclaw agent --help
  • openclaw skills --help
• Do not require OpenClaw or Hermes to be installed for normal Codex interactive use. Missing runtimes should degrade to a chat report.

Safety#

• Do not hand edit weight, last_active, hot, or derived dashboard values.
• Do not write to links/; that layer is retired.
• Do not log anchor usage for AI-only analysis.
• Ask before deleting, renaming, merging, retiring, migrating, or editing raw source material.
• Writing a report under _meta/anchor-governance/knowledge-growth/ is allowed when the user asks for a persistent report; otherwise answer in chat.