Skip to content

agent-playbook.md

Latest commit

 

History

History
99 lines (69 loc) · 2.44 KB

agent-playbook.md

File metadata and controls

99 lines (69 loc) · 2.44 KB

OpenCLI Agent Playbook

This file is optimized for agent behavior. Read it when the user wants results through OpenCLI and you need a safe default path quickly.

Trigger Phrases

Strong signals:

  • "use opencli"
  • "check what opencli can do"
  • "run this through my browser login"
  • "query Bilibili / Zhihu / Xiaohongshu / Weixin / Boss"
  • "control Codex / Cursor / ChatGPT / Notion from terminal"
  • "download this article/video with opencli"
  • "generate an opencli adapter"

Weak but useful signals:

  • The task needs deterministic CLI output instead of free-form browsing
  • The user already has the site logged in in Chrome
  • The task targets a known site repeatedly

Default Action Order

Follow this order unless the user already gave an exact command:

  1. opencli doctor if the task is browser-backed and health is unknown
  2. opencli list -f yaml if the command surface is unknown
  3. Use read-only commands first
  4. Use -f json if another tool or model will consume the result
  5. Expand scope only after a small first pass succeeds

First Commands By Situation

Browser-backed site

opencli doctor
opencli list -f yaml
opencli <site> <read-command> --limit 5 -f json

Desktop adapter

opencli <app> --help
opencli <app> read -f json
opencli <app> history -f json
opencli <app> model -f json

External CLI passthrough

opencli <tool> ...

Adapter development

opencli explore <url> --site <name>
opencli synthesize <name>
opencli generate <url> --goal "<goal>"

Safe Defaults

  • Start with --limit 5 or --limit 10
  • Prefer json for machine consumption
  • Prefer yaml when the human will scan raw output
  • Do not begin with write actions such as post, reply, publish, greet, send, or follow
  • For browser-backed commands, suspect login state before suspecting the adapter

Write-Action Rule

Before any write-capable command:

  1. Read current state first
  2. Confirm target identity or content source
  3. Keep the outbound command explicit in the transcript

Examples:

  • Before boss greet, run boss detail
  • Before jike post, inspect timeline or notifications if relevant
  • Before desktop send or ask, confirm the app and active context with read, history, or model

Installed-Version Rule

Do not assume the README equals the installed surface area.

When in doubt:

opencli <adapter> --help
opencli list -f yaml

Trust the installed CLI over stale docs.