clawtouch-skills
June 4, 2026 · View on GitHub
English | 简体中文
clawtouch-skills
Markdown skill files for LLM agents driving real-world apps through HID. Concise operator manuals for specific applications — what keyboard shortcut, what menu path, what gotcha — written for an LLM that's about to use
clawtouch-mcp(or any compatible HID stack) to drive that application.
🌐 clawtouch.cn — official site for hardware, docs, and commercial inquiries.
What's a skill?
A skill is not code. It's a markdown file you load into the LLM's context before a task. It tells the model things it might not know on its own — exact menu paths, keyboard shortcuts, version differences, quirks that bite agents but a human would never notice.
The model still decides what to do; the skill just gives it the local knowledge to do it well.
Why these specific apps?
LLM training data is heavy on US/EN-language apps (Word, Excel, Photoshop, …) and thin on Chinese-market apps (WPS, Feishu / Lark, DingTalk). The skills here cover the second group, where the delta between "LLM guesses" and "actual UI" is widest — and the value of a skill file is highest.
Most skills are per-app operator manuals; a few are cross-cutting techniques (e.g. reliable text entry) that apply across all of them.
PRs adding skills for other apps welcome — see CONTRIBUTING.md.
Available skills
Each skill lives in skills/<skill-name>/SKILL.md — a single markdown
file with a YAML frontmatter block (name + description) at the top,
matching the Anthropic Claude Skills
and GitHub Copilot Agent Skills
conventions, so any compatible client can discover them.
| Skill | App | Coverage |
|---|---|---|
skills/wps-office/SKILL.md | WPS Office (文字 / 表格 / 演示) | Common editing tasks; shortcut differences from MS Office |
skills/feishu/SKILL.md | Feishu / Lark (飞书) | IM, docs, sheets, calendar |
skills/dingtalk/SKILL.md | DingTalk (钉钉) | IM, file sharing, docs, calendar |
skills/paste-text/SKILL.md | All apps (technique) | Reliable non-ASCII / Chinese text entry via clipboard paste (works around IME & keyboard-layout issues) |
skills/grid-ui-automation/SKILL.md | All apps (technique) | Positional clicking on a regular grid — calibrate from a screenshot, batch the clicks, verify & retry (minesweeper, keypad, spreadsheet cells, tile puzzles) |
How to use
- Pick the skill matching the application you're about to drive.
- Load it into your LLM context. For
Claude Code,
drop the
skills/<name>/directory under~/.claude/skills/(user-scoped) or.claude/skills/(project-scoped) and Claude Code will auto-discover it via the frontmatter. For other clients (Cline, Continue, OpenClaw, custom agents), either install via aSKILL.mdregistry tool or include the file contents in the initial system prompt. - Make sure
clawtouch-mcpis connected and--screen WxHcovers the right monitor. Skills that depend on drag or held-key gestures (hid.drag,hid.mouse_button_down/up,hid.key_press/release,hid.hold_key) requireclawtouch-mcp v0.3.0+for protocol v1.1 support; skills using onlyhid.click,hid.type,hid.scroll,hid.keyetc. work with any release. - Issue the high-level task — the model uses the skill's knowledge to pick the right HID calls.
Scope — documentation, not content generation
Skill files are operator manuals written for an LLM. They are markdown text documents only — no executable code, no AI model, no content-generation logic. The calling LLM agent is solely responsible for any text, image, audio, or video content it produces while following the skill, including compliance with any content-labeling or content-moderation obligations applicable in the agent's jurisdiction (e.g. PRC AI Generated Content Labeling Measure effective 2025-09-01).
Acceptable use
These skill files describe legitimate uses of the target applications — accessibility tooling, internal RPA on accounts the operator owns, QA / compatibility testing, personal automation. This project does not support, document, or assist with use cases that:
- Bypass, evade, or interfere with any application's anti-fraud, anti-abuse, rate-limiting, or risk-control measures.
- Operate accounts the user does not lawfully own or have explicit authorization to operate.
- Are prohibited by the target application's Terms of Service in the user's jurisdiction (check the ToS before applying any skill to a specific use case).
- Violate applicable law — including, but not limited to, PRC Anti-Unfair Competition Law Art. 13 (the Internet sector specific provision; promulgated 2025-06-27, effective 2025-10-15) covering improper means — including circumventing technical management measures — to acquire or use another operator's data; Personal Information Protection Law; Cybersecurity Law; and equivalent laws in other jurisdictions.
These statements describe the scope of our maintainer support and documentation — they are not additional restrictions on the MIT License, which continues to govern all use, modification, and redistribution of the markdown text content. Users are independently responsible for evaluating their specific use case against applicable laws and the target application's ToS. Contributors are required to flag any skill content that risks crossing these lines (see CONTRIBUTING.md).
Soft guidance, not enforcement
Skills are soft guidance — the LLM still decides what to do, and
might not follow the skill exactly. That's fine for development,
research, or personal automation. Stricter execution lives in the
closed-source ClawTouch desktop product (contact support@tinqiao.com).
Limitations
- Skills describe UIs as of early 2026. Apps update; PRs welcome when you notice drift.
- Hardcoded screen coordinates are intentionally absent — UI layouts
differ by window size, DPI, theme, locale. Skills name keyboard paths
whenever possible, since those are stable across layouts. When you must
click by position (a regular grid with no keyboard route), the
grid-ui-automationtechnique derives coordinates from a live screenshot at runtime instead of baking them in. - IME / input-method state is not handled here — set the host IME to the right state before invoking literal text typing.
Naming convention
Each skill is a directory at skills/<skill-name>/ containing a
SKILL.md file. <skill-name> is all-lowercase kebab-case (e.g.
wps-office, feishu, dingtalk) and must match the name: field
in the SKILL.md frontmatter.
One skill per directory. If an app has very different sub-products
(Office's Word vs Excel) where one bundled file would be too long for
useful agent loading, split into skills/word/SKILL.md /
skills/excel/SKILL.md rather than one giant skill.
For Chinese-localized apps, the directory name uses the romanized name
(wps-office, feishu, dingtalk); the Chinese name goes in the
SKILL.md title and first sentence.
Companion repositories
| Repo | Role |
|---|---|
clawtouch-mcp | MCP server exposing HID tools — drives the actual hardware |
clawtouch-hid | Pico 2 firmware + v1.1 wire protocol (v1.0 baseline frozen) |
| (this repo) | Markdown skill files for app-specific guidance |
Related work
The agent-skill format ecosystem is converging on YAML-frontmatter
markdown bundles. clawtouch-skills adopts the
Anthropic Claude Skills
layout — skills/<skill-name>/SKILL.md with name / description
frontmatter — so skills here can be loaded by any compatible client
without conversion.
- Anthropic Claude Skills — the format we follow.
- GitHub Copilot Agent Skills — same frontmatter convention; targets Copilot CLI and cloud agents.
- Community CLIs such as
kcchien/skills-cliinstallSKILL.mdbundles into Claude / Cline / Continue registries.
What's different about ClawTouch skills:
- HID-mode skills, not API skills. Most published skill packages
teach an LLM to call a service API
(OpenClaw skill packages,
GitHub / Slack / Notion connectors). ClawTouch skills teach the LLM
how to drive the UI of those apps via simulated keyboard / mouse —
the path that's available when there is no API (WPS Office desktop,
Feishu / DingTalk native clients, internal enterprise tools without
REST endpoints) — through
clawtouch-mcp. - Chinese-market app focus. LLMs are well-trained on Word / Slack / Notion / Google Docs but have far less ground truth on WPS / Feishu / DingTalk. These skills target the gap where explicit operator knowledge has the highest marginal value.
Contributing
We welcome new skills, corrections to existing ones, and translations. See CONTRIBUTING.md for the file structure convention and PR checklist.
About
clawtouch-skills is maintained by Tinqiao Technology — the
team behind ClawTouch (clawtouch.cn),
building plug-in USB devices that let LLM agents operate real
Windows / macOS / Linux desktops at the HID layer.
License
MIT © Tinqiao Technology (Beijing) Co., Ltd. — see LICENSE (English, authoritative) and LICENSE.zh-CN.md (non-official Chinese translation, for reference).
This repository contains no bundled third-party source code — see NOTICE. Trademarks (ClawTouch, Tinqiao, and the third-party application marks named in skill files — WPS Office, Feishu / Lark, DingTalk) are covered separately in TRADEMARKS.md — the MIT License does not grant any trademark rights and no third-party trademark owner has endorsed or sponsored this documentation.
For commercial deployments at scale, enterprise support, or OEM
hardware discussion: support@tinqiao.com.