Tashan

Learn best practices, extract the working protocol, and turn external examples into your own Prompt / SOP / Skill.

中文 · Install Guide

Tashan is a methodology skill for AI agents. It helps you study an unfamiliar domain, compare strong repositories/tools/skills/products, and turn what you learn into your own workflow, prompt, SOP, or skill.

The core idea:

Stones from other hills may polish jade.

In plain language: learn from strong outside examples to sharpen your own working system.

When to use it

Use Tashan when:

• You find a GitHub repository and want to know whether it can be useful to you.
• You are entering an unfamiliar domain and want to learn from best practices first.
• You want to compare multiple tools instead of worshipping one project.
• You want to turn outside experience into your own Prompt, SOP, Agent, or Skill.
• You explicitly do not want to build immediately; you want to study, deconstruct, and synthesize first.

What it does

Tashan turns learning into a 10-step protocol:

1. Define the boundary: learning, review, benchmarking, or execution.
2. Find references: collect 3-10 strong examples and useful counterexamples.
3. Build an index: source, local path, version, value, and what not to copy.
4. Extract the object model: how experts make the task operable.
5. Extract the working protocol: who decides, who executes, who reviews.
6. Extract quality gates: planning gates, intermediate artifacts, delivery checks, review loops.
7. Extract hidden assumptions: what the tool believes humans and machines should each do.
8. Compare horizontally: common principles, conflicts, and transferable patterns.
9. Distill invariants: rules that remain true across implementations.
10. Rebuild your own system: Prompt / SOP / Skill draft.

Quick install

One-line install for Codex

mkdir -p ~/.codex/skills/他山 && curl -fsSL https://raw.githubusercontent.com/zhenxishuai/tashan-skill/main/SKILL.md -o ~/.codex/skills/他山/SKILL.md

Install from a local checkout for Codex

mkdir -p ~/.codex/skills/他山
cp SKILL.md ~/.codex/skills/他山/SKILL.md

From this project:

mkdir -p ~/.codex/skills/他山
cp skills/他山/SKILL.md ~/.codex/skills/他山/SKILL.md

One-line install for Claude Code

mkdir -p ~/.claude/skills/他山 && curl -fsSL https://raw.githubusercontent.com/zhenxishuai/tashan-skill/main/SKILL.md -o ~/.claude/skills/他山/SKILL.md

Install from a local checkout for Claude Code

mkdir -p ~/.claude/skills/他山
cp SKILL.md ~/.claude/skills/他山/SKILL.md

From this project:

mkdir -p ~/.claude/skills/他山
cp skills/他山/SKILL.md ~/.claude/skills/他山/SKILL.md

Restart your agent session after installation so the skill list is reloaded.

For detailed options, see INSTALL.md.

Usage

Example prompts:

Use Tashan to study how AI investment research reports are built. Find strong repositories and products first. Do not build yet.

This GitHub repository looks useful. Review whether we can learn from it and what should not be copied.

I know nothing about AI sales follow-up. Learn from best practices and turn the result into a skill.

Expected outputs

By default, Tashan produces three types of files:

references/[topic]/README.md              # reference index
docs/YYYY-MM-DD-[topic]-tashan-notes.md     # comparative learning notes
skills/[name]/SKILL.md                    # synthesized skill draft

Depending on the project, it may also produce:

docs/[topic]_PROMPT_SPEC.md
docs/[topic]_SOP.md
docs/[topic]_BENCHMARK.md

Why it is not ordinary research

Ordinary research often stops at “what features does this tool have?” Tashan goes deeper:

• How does it define the task?
• What object model does it use?
• What should the machine do, and what should the human decide?
• How does it prevent low-quality output?
• How does it turn feedback into better future rules?
• What should be transferred, and what should not be copied?

Project structure

他山/
├── SKILL.md              # skill body
├── README.zh-CN.md       # Chinese README
├── README.md             # English README
├── INSTALL.zh-CN.md      # Chinese install guide
├── INSTALL.md            # English install guide
└── evals/
    └── evals.json        # example evaluations

Design principles

• Do not rush into using a tool; understand the problem first.
• Do not learn from one sample only; compare multiple references.
• Do not only learn features; extract object models, protocols, and quality gates.
• Do not confuse tool capability with domain capability.
• Do not let the main executor be the final reviewer.
• Save the learning result into files; do not leave it only in chat.

License

MIT is recommended. If maintained as a long-term open-source project, add CONTRIBUTING.md and release notes.