AI Agent Skills for Chinese Knowledge Workers

Three Kings 2.2: compress with iMandalArt, address with BIRD, and publish with an A4 eight-page booklet.

iMandalArt, FIRE semantic analysis, planning, and publishing workflows for Claude Code, Codex, and mainstream LLM agents.

This repository is TWHSI's public agent-skill system for Chinese knowledge workers. It turns repeatable writing, planning, note-making, and publishing workflows into skills that can be used by Claude Code, Codex, ChatGPT, Gemini, Hermes, and other LLM agents.

The source of truth lives in skills/*/SKILL.md; skill-tree.json records the Three Kings relationship, and the website turns the skills into generated LLM manifests, a searchable index, install commands, revisions, and update timestamps.

What This Repository Is

twhsi/skills is an installable, shareable, LLM-readable skill registry. It turns long-running workflows for weekly review, iMandalArt, FIRE analysis, card-based notes, EPUB publishing, HyperCard returns, and desktop work into public procedures that agents can repeat and humans can inspect.

The point is not to make humans behave like faster machines. The point is to let AI spend tokens on repetitive structure, while humans keep judgment, relationships, pace, and the quieter question:

Which action brings real happiness and peace?

The flagship workflow is Skills 三大天王 2.2: imandalart compresses a source, thebrain-bird-address gives every knowledge object an address and Route, and a4-eight-page-booklet returns the structure to a verified paper book.

Live site: https://www.twhsi.com/
Live Keyword Graph: https://keyword-graph-view.twhsi.chatgpt.site/

Agent manifest: https://www.twhsi.com/agent.json
Skill index: https://www.twhsi.com/skills.json
LLM context: https://www.twhsi.com/llms.txt

What It Does

• Convert source material into iMandalArt 2.2 hard-line 9-grid cards.
• Apply FIRE semantic analysis for card-box thinking, retrieval, and writing structure.
• Turn permanent notes into BIRD 2.1 Book Addresses, structured Knowledge Indexes, Routes, verified Deep Links, Semantic Roles, and TheBrain manuscript scaffolds.
• Compress answers with concise-key-points: reduce default length by 70% while retaining at least 95% of essential information.
• Compile source material into a verified A4 one-sheet eight-page booklet, synchronized reading PDF, print imposition, and editable DOCX.
• Generate agent-readable metadata through agent.json, skills.json, and llms.txt.
• Route daily planning, weekly planning, booklet, EPUB, Markdown, and knowledge-management workflows.
• Extract eight keywords and render a distributed weighted graph with definitions, evidence, and blacklist control.
• Package Chinese-first workflows so LLM agents can operate on them consistently.

Copyable Demo

Input: manuscript, notes, PDF, or long conversation
Skills: imandalart 2.2 -> thebrain-bird-address 2.2 -> a4-eight-page-booklet 2.2
Output: eight-angle index -> BIRD addresses and Routes -> reading PDF + print sheet + DOCX

Skills 三大天王 2.2

素材
  -> iMandalArt 2.2：一個中心、八個正交視角
  -> BIRD 拆書分析 2.2：Book Address、W+T+K+A、Route、Deep Link
  -> A4 八頁小書 2.2：八頁閱讀版、折頁列印版、可編輯 DOCX、驗證報告

2.2 New Functions

• Shared handoff：中心、八角度、來源 ID、BIRD Address、Route 與引用一路保留到紙本。
• Seven-Route publishing：第 1 頁放 iMandalArt 總覽，第 2–8 頁放七條 BIRD Routes。
• Backward compatibility：BIRD 仍輸出 BIRD-2.1；小書仍讀取 BookletManifest 2.0。
• Machine-readable tree：skill-tree.json 記錄三者的版本、成熟度、輸出與解鎖關係。

Featured: Keyword Graph View

keyword-graph-view turns Chinese, English, or mixed text into a distributed network of exactly eight keywords. It removes blacklisted terms, weights co-occurrence edges from W1 to W9, and gives every node a definition, note, and source evidence.

Open the live tool: keyword-graph-view.twhsi.chatgpt.site

Featured: Weekly Reverse Review

weekly-reverse-review turns an 11-step weekly review into an AI-assisted reverse planning loop.

The skill reads four angles before writing the next week:

• YEAR: annual plan and hundred-year life plan.
• Week: last weekly plan and weekly review.
• Day: daily plans, seven-day diary notes, and calendar evidence.
• Inbox: loose tasks, reminders, subscriptions, errands, and collected noise.

It then asks what should be smaller, slower, deleted, delayed, or kept as presence instead of achievement. The center question is:

Which action brings real happiness and peace?

Default answer:

Speak less. Stay present. Move slowly.

For an LLM or agent, start with:

curl -s https://www.twhsi.com/llms.txt

Deep Dive: iMandalArt 2.2

imandalart turns loose source material into one hard-line 3x3 Mandala index card.

iMandalArt 2.2 is designed around a strict text contract:

• Eight orthogonal surrounding angles labeled Ⓐ through Ⓗ.
• A double center axis displayed as ◎◎◎◎◎.
• Exactly 11 physical text lines, so the card survives chat previews, note apps, and clipboard workflows.
• Compact CJK-friendly cells for TheBrain/Cerebro, Hermes, Discord, Codex, and other LLM chat surfaces.

Current status: iMandalArt 2.2 is optimized for CJK workflows, pure-text stability, の字型 identity preservation, and clean handoff to BIRD and booklet Routes.

Conceptual example:

ⒶHealth Reset  ⒷManuscript  ⒸMoney Noise
Move with care  Pull one center  Review first
Build rhythm    Draft before all Fund the work

ⒹFamily Spark  ◎◎◎◎◎    ⒺPeople Path
Ask not lecture Weekly Review Keep three cards
Walk the long   ◎◎◎◎◎    Connect next

ⒻInner Release ⒼLearning    ⒽJoyful Rest
Speak less      Number lines  Sing and loosen
Prove nothing   Build index   Guitar and sun

Use it for weekly planning, writing focus, knowledge capture, and CJK note workflows where visual stability matters as much as semantic compression.

Deep Dive: BIRD and A4 2.2

BIRD Book Deconstructor 2.2

thebrain-bird-address turns manuscripts and permanent notes into addressable knowledge objects. The 2.2 Skill release applies the backward-compatible BIRD 2.1 protocol: B = Book Address, I = Weight + Type + Keyword + Alias, R = Route, and D = exact Deep Link, with optional evidence-based Semantic Roles. It supports Excel, TheBrain scaffolds, Roam JSON, monochrome double nine-grids, and a direct seven-Route booklet handoff.

Supporting Skill: 言簡意賅｜講重點｜Concise

concise-key-points produces high-information-density answers: conclusion first, three to five key points, minimal repetition, and no unnecessary background. Its default target is 70% shorter while retaining at least 95% of essential information.

A4 Eight Page Booklet 2.2

a4-eight-page-booklet turns text, notes, images, BIRD Graph JSON, iMandalArt grids, 64+8+1 structures, FIRE material, daily plans, Bike or Bonsai outlines, templates, and PDFs into a printable, cuttable, and foldable eight-page booklet made from one A4 sheet.

The skill does more than place content into eight panels. It separates content from layout, normalizes the source into BookletManifest 2.0, and generates every reading, printing, preview, and validation artifact from the same manifest.

• Page 1 can hold the cover and overview. In BIRD or 64+8+1 mode, it contains one directory nine-grid and one keyword nine-grid.
• Pages 2–8 can form seven Routes. Each page keeps one central claim, three to five concise points, and one ordered reading path.
• iMandalArt spiral-order coordinates are preserved together with source IDs, BIRD Addresses, citations, and cross-route notes.
• Outputs include eight upright A7 page PDFs, an eight-page reading PDF, an A4 landscape print imposition, an editable DOCX, preview images, and a validation report.
• The fixed imposition order is 6, 7, 8, 1 upright on the top row and 5, 4, 3, 2 rotated 180 degrees on the bottom row.
• Validation checks structure, typography, imposition, visual quality, Apple Pages compatibility, and source fidelity when required.

Full skill specification and installation source: skills/a4-eight-page-booklet/SKILL.md

What This Is

• A GitHub-backed skill registry for Chinese agent workflows.
• A machine-readable index for LLM agents that need routing, install commands, and resource discovery.
• A human-readable map for deciding which workflows should become reusable skills.
• A static website that publishes each skill's semantic version when declared, latest Git revision, and latest update time.

Core Skill Stack

Current Metadata Highlight

fire-analysis-card is now on FIRE 2.0:

• F = Full-D: stable numbering for temporary, permanent, and project notes.
• I = Index: keyword webs and retrieval handles.
• R = Route: thinking paths through the material.
• E = Evolution: time-based card-box growth for semantic search.

Skill Versions And Updates

The live website generates freshness metadata from skill files and Git history on every deployment. Each skill entry includes:

• version: the semantic skill version when declared, otherwise Unversioned
• revision and revision_short: the latest Git commit for that skill path
• updated_at: the latest update timestamp for that skill path

Open the live generated table:

• https://www.twhsi.com/#updates
• https://www.twhsi.com/skills.json

Featured release check: iMandalArt, BIRD Book Deconstructor, and A4 Eight Page Booklet are presented together as Skills 三大天王 2.2. Their stable interchange formats remain backward compatible.

thebrain-bird-address implements BIRD 2.1: B = Book Address, I = Weight + Type + Keyword + Alias, R = Route, and D = exact Deep Link, with optional evidence-based Semantic Roles. It complements FIRE by moving from reusable permanent notes into project notes, chapter structures, and book production.

Registry Routes

Core Files

assets/      public display assets
skills/      installable skills for mainstream LLM workflows
docs/        install notes, book links, and skill index
examples/    sample inputs and outputs
archive/     older drafts and retired skills
Hermes.md    system map and command file
skill-tree.json  Three Kings 2.2 machine-readable relationship map
site/        static website source
dist/        generated website output, ignored by git

Install A Skill Locally

From the repository root:

cp -R skills/fire-analysis-card ~/.codex/skills/
cp -R skills/thebrain-bird-address ~/.codex/skills/
cp -R skills/concise-key-points ~/.codex/skills/
cp -R skills/todays-daily-plan ~/.codex/skills/
cp -R skills/weekly-reverse-review ~/.codex/skills/
cp -R skills/imandalart ~/.codex/skills/
cp -R skills/keyword-graph-view ~/.codex/skills/
cp -R skills/a4-eight-page-booklet ~/.codex/skills/

Then validate a skill:

python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py skills/fire-analysis-card

Build The Website

npm run build

The build reads skills/*/SKILL.md, extracts frontmatter, detects declared versions, asks Git for latest per-skill revision timestamps, and writes:

• dist/agent.json
• dist/skills.json
• dist/llms.txt
• the static HTML/CSS/JS site

Vercel configuration lives in vercel.json: build command npm run build, output directory dist.

Maintenance Rule

When a skill changes, update the skill file first, run validation, rebuild the website, and push to main. The live site should always show the latest skill version, revision, and update time.