AppClaw Agent Skill

A cross-agent skill for driving and inspecting mobile apps interactively through the appclaw-agent CLI — open apps, press/fill UI elements, scroll, and verify on-screen state with screenshots, all without writing a YAML flow.

Works with Claude Code, Gemini CLI, and OpenAI Codex CLI. The behavioral instructions are identical across all three; only the install location and invocation differ.

Prerequisites

The skill drives the appclaw-agent binary. Install it first:

npm install -g appclaw-agent
appclaw-agent help workflow   # sanity check

appclaw-agent needs an Appium-backed device/emulator session, as documented in the AppClaw project.

---

Claude Code

Plugin marketplace (recommended)

/plugin marketplace add agentfarmhq/appclaw-agent-skill
/plugin install appclaw@appclaw-agent-skill

Invoke with /appclaw:use-appclaw-agent-cli, or just ask Claude to "open app X and tap Y" — it auto-triggers from the skill description. Manage with /plugin list, /plugin disable appclaw@appclaw-agent-skill, etc.

Manual install

cp -r skills/use-appclaw-agent-cli ~/.claude/skills/        # personal
# or:  cp -r skills/use-appclaw-agent-cli .claude/skills/   # project-scoped

Invoke with /use-appclaw-agent-cli.

---

OpenAI Codex CLI

Codex uses Agent Skills — the same SKILL.md (with name/description frontmatter) as Claude Code. Copy the skill folder into the Codex skills dir:

mkdir -p ~/.agents/skills
cp -r skills/use-appclaw-agent-cli ~/.agents/skills/        # user-wide
# or:  cp -r skills/use-appclaw-agent-cli .agents/skills/   # repo-scoped

Restart Codex. It triggers implicitly from the description, or run it explicitly via the /skills menu or $use-appclaw-agent-cli.

Note: Codex's older ~/.codex/prompts/ custom prompts are deprecated — Agent Skills are the supported path.

---

Gemini CLI

Gemini uses an extension (this repo's root gemini-extension.json + GEMINI.md). Install straight from git:

gemini extensions install https://github.com/agentfarmhq/appclaw-agent-skill

Once installed, the GEMINI.md context is loaded automatically, so just ask Gemini to operate the app. Manage with:

gemini extensions list
gemini extensions disable use-appclaw-agent-cli
gemini extensions uninstall use-appclaw-agent-cli
gemini extensions update use-appclaw-agent-cli

Prefer no install step? Drop the context file in manually:

cp GEMINI.md ~/.gemini/GEMINI.md        # global, all projects
# or:  cp GEMINI.md ./GEMINI.md         # project root

---

What it does

Once active, the skill guides the agent to:

• Open a named device session and inspect actionable UI (snapshot -i --json)
• Interact via stable @eN refs / selectors (press, fill, longpress, scroll)
• Always use scroll (never the ambiguous swipe); the tap verb is press, not tap
• Verify assertions visually via screenshots, never DOM presence alone
• Close the session when done

The canonical instructions live in skills/use-appclaw-agent-cli/SKILL.md (Claude + Codex). GEMINI.md is generated from it — edit SKILL.md, then run:

scripts/sync-gemini.sh           # regenerate GEMINI.md
scripts/sync-gemini.sh --check   # CI guard: fail if GEMINI.md is stale

Repository layout

appclaw-agent-skill/
├── .claude-plugin/            # Claude Code plugin + marketplace
│   ├── plugin.json
│   └── marketplace.json
├── skills/
│   └── use-appclaw-agent-cli/
│       └── SKILL.md           # Claude Code skill + Codex Agent Skill
├── gemini-extension.json      # Gemini CLI extension manifest
├── GEMINI.md                  # Gemini CLI context (generated from SKILL.md)
├── scripts/
│   └── sync-gemini.sh         # regenerate GEMINI.md from SKILL.md
└── README.md

License

MIT