# Qiushi Skill: 'Seeking Truth' Skills for AI Agents
Languages:
简体中文 |
English
> Build agents that investigate first, focus on the main contradiction, validate in practice, and keep pushing until the work is actually done.
`qiushi-skill` is a collection of reusable methodology skills, commands, and session-start hooks. It turns a set of problem-solving methods into installable assets for Claude Code, Cursor, Codex, OpenCode, OpenClaw, Hermes Agent, and other prompt-driven hosts.
## Why This Exists
Many agents can generate text, but they still fail at disciplined work:
- they answer before collecting facts
- they scatter effort across too many tasks
- they stop at the first blocker
- they skip verification and call that "done"
`qiushi-skill` is meant to correct that with one core principle and nine concrete methods:
- `seeking truth from facts` as the governing principle
- contradiction analysis
- practice -> cognition -> practice iteration
- investigation first
- mass line feedback synthesis
- criticism and self-criticism
- protracted strategy
- concentrate forces
- spark a prairie fire
- overall planning
> Support author's other repo!: **[AcademicForge](https://github.com/HughYau/AcademicForge)**: **One Forge, All Skills.** An interactive, cross-platform tool to generate simple, one-click installation commands for multiple AI Agent Skills.
## Install
### Preferred: `npx qiushi-skill`
```bash
npx qiushi-skill
```
Interactive by default: it detects installed hosts and copies into the right directories. Non-interactive:
```bash
npx qiushi-skill install --target claude-code --scope user
npx qiushi-skill install --target all --scope user
npx qiushi-skill uninstall --target claude-code
```
The CLI is zero-dependency, writes a `.qiushi-skill-install.json` manifest per target so uninstall only removes what it installed, and ships `validate` for self-checks.
### Claude Code Official Marketplace
This repository now ships `.claude-plugin/marketplace.json`, so Claude Code can discover it directly from GitHub:
```text
/plugin marketplace add HughYau/qiushi-skill
/plugin install qiushi-skill@qiushi-skill
```
### Source Install
```bash
git clone https://github.com/HughYau/qiushi-skill
cd qiushi-skill
claude --plugin-dir .
```
### Other Hosts
The same CLI targets Cursor, Codex, OpenCode, OpenClaw, Hermes Agent, and nanobot (`--target cursor|codex|opencode|openclaw|hermes|nanobot|all`). Any host that reads Markdown skills can also just copy the `skills/` directory. Target paths and native entrypoints are listed in [`docs/platforms.md`](https://github.com/HughYau/qiushi-skill/blob/main/docs/platforms.md).
## Validate
Preferred:
```bash
npx qiushi-skill validate
```
If you are working from a source checkout, legacy fallback scripts are also available:
```bash
bash tests/validate.sh
# or on Windows
powershell -NoLogo -NoProfile -ExecutionPolicy Bypass -File tests/validate.ps1
```
Validation checks:
- JSON validity, including `.claude-plugin/marketplace.json`
- version consistency between `package.json` and marketplace metadata
- presence of CLI files, hooks, commands, and the marketplace bundle
- frontmatter completeness for skills, commands, and agents
- local Markdown link integrity
## How to Use It
The session-start entry skill is `skills/arming-thought/SKILL.md`, a ~50-line always-on kernel. Its job is intentionally narrow:
1. enforce `seeking truth from facts` through four hard rules: claims follow evidence, separate fact / inference / unknown, verified means done, diagnose before giving up
2. route to a downstream skill only when it clearly helps; direct-execution tasks trigger nothing
3. defer to the host's own plan / review / todo flows when they exist
Every method skill keeps only five sections: **use / don't use** (trigger boundary), **procedure** (concrete actions), **output template** (an observable artifact), **discipline** (hard constraints), and **handoff** (which skill usually comes next). Original quotations and lookup guides live in sibling files and are loaded on demand.
Two host-discoverable sub-agents live in `agents/`: `investigator` (read-only, returns a fact / inference / unknown ledger) and `self-critic` (fresh-context reviewer that reads artifacts, not narratives).
Manual command entrypoints live in `commands/*.md` for hosts that support Markdown slash commands.
## Project Layout
```text
qiushi-skill/
├── .claude-plugin/
│ ├── marketplace.json
│ └── plugin.json
├── .cursor-plugin/plugin.json
├── bin/ # npx qiushi-skill CLI
├── skills/ # entry kernel + nine method skills + workflows
├── commands/ # manual slash-command entrypoints
├── hooks/ # SessionStart injection (POSIX + PowerShell)
├── agents/ # investigator, self-critic
├── docs/ # site page + platforms.md
├── CHANGELOG.md
├── README.md
└── README.en.md
```
## License
MIT