---

name: agent-cli-creator description: Use when the user wants to build a CLI tool to automate browser interactions on a specific website using kimi-webbridge. Invoke when user says "create a CLI for X site", "build a tool to automate X", or wants to control a website programmatically via an AI agent.

Agent CLI Creator

Guide an AI Coding Agent to build a website-automating CLI tool backed by the kimi-webbridge browser daemon.

Phase 1: Prerequisites

~/.kimi-webbridge/bin/kimi-webbridge status

| Result | Action | |--------|--------| | running: true + extension_connected: true | Proceed | | Command not found | Tell user to visit https://www.kimi.com/features/webbridge (中文: https://www.kimi.com/zh-cn/features/webbridge) to install | | running: false or extension_connected: false | Invoke the kimi-webbridge skill → references/operations.md |

Phase 2: Requirements Interview

Ask the user in one message, wait for reply:

1. Target website URL (required)
2. Programming language — Go (recommended; references/go-layout.md is the canonical template) / Python / Node.js / Other
3. Login required? — Yes / No / Unknown (can skip for now)
4. First 1–3 features — pick from common categories:
   • Read: home feed, search, profile page, post/item detail
   • Write: create post, like/unlike, comment, bookmark/save
   • Account: login-status, user info

Explain iterative development to the user:

"We'll start with 1–3 features to validate the approach end-to-end before adding more. Site exploration for features you don't need yet wastes time. You can always add features later by re-running from Phase 3."

Phase 3: Site Exploration (mandatory before writing any code)

This phase is non-negotiable. Do not write business logic until exploration is complete.

For each planned feature, run the full protocol in references/site-exploration.md.

The protocol yields: API endpoint, required headers, response shape, and a verified evaluate call that proves the API works inside the browser session.

Sanity-check every limitation before you write it down: would a human clicking through the same flow in their own browser hit it too? If not, the difference is in your automation setup — tab visibility, event timing, hydration — not in the site. A real one: "search caps out at 20 results", when a human scrolling that page gets 200 — see references/site-exploration.md → The Background Tab Trap.

Only proceed to Phase 4 when you have a working evaluate call for every planned feature.

Phase 4: Implement

Order matters — do not skip ahead:

1. Project scaffold — see references/go-layout.md for Go; adapt conventions for other languages
2. login-status command — if the site requires login; see references/login-handling.md
3. Read commands — no side effects; implement and test first
4. Write commands — side effects (post, like, etc.); implement after reads work

After each command is implemented, immediately verify before moving on:

{platform}-cli {command} --help          # --help must work
{platform}-cli {command} [args]          # must return {"ok": true, "data": ...}

For write commands, also verify the error path (e.g., wrong ID, missing flag) returns {"ok": false, ...} with non-zero exit.

If verification fails, fix before implementing the next command.

Universal CLI contract (all languages):

• --help / -h must work on every command
• All output: {"ok": true, "data": ...} or {"ok": false, "error": {"code": "...", "message": "..."}}
• Non-zero exit code on error

Go: 用 go mod init {platform}-cli 初始化独立 module，按 references/go-layout.md 的结构搭建，包含自己的轻量 browser client 和 output helper。

Phase 5: Write Companion Skill

After the CLI works, create ~/.claude/skills/{platform}-cli/SKILL.md using the template in references/companion-skill-template.md.

Purpose: tells a future AI agent how to use the CLI, not how to build it. The two skills serve different audiences at different times.