social-push: A Claude Code Skill That Drafts Social Media Posts via Browser Automation

What social-push Does and Who It Is For

Social media posting involves repetitive browser interactions: navigate to the platform, fill the content fields, attach images, and submit. For developers who publish technical articles, project announcements, or product updates to multiple platforms, doing this manually across five or six sites is time-consuming. The social-push skill brings this workflow into Claude Code: you describe the content you want to post in natural language and invoke the skill, and the agent handles the browser interactions.

The skill targets developers and content creators who maintain a presence across both Chinese platforms (Xiaohongshu, Weibo, Zhihu, WeChat public accounts, Juejin, Linux.do) and international ones (X/Twitter). It is built with Claude Code's skill system and relies on agent-browser for the actual browser automation.

agent-browser Versus Playwright MCP

Most browser automation approaches for AI agents use Playwright MCP, which passes raw DOM snapshots and interaction references to the model. The README notes two issues with that approach: it consumes large numbers of tokens per interaction, and it is slow because of the round-trip between DOM snapshot, model response, and browser action.

agent-browser takes a different approach. It parses interactive references from the page and reduces what needs to be sent to the model, lowering token consumption and improving speed. The README specifically describes running agent-browser from bash with --help to get well-formed instructions for each step, which makes the interaction pattern faster than full DOM exploration.

Using agent-browser also means the skill benefits from Claude Code's bash integration directly. Commands like agent-browser connect, agent-browser navigate, and state management via --state or --profile flags are invoked as shell commands within the agent session, not through a separate MCP server managing browser sessions.

Self-Evolution and Markdown-as-Configuration

Social media platform layouts change frequently. A selector-based automation script that works today may break after a platform redesign. The social-push skill uses a design it calls self-evolution: when the expected page structure is not found, the skill can detect the change and attempt to repair its own workflow file before retrying.

This is possible because each platform's publishing workflow is encoded in a separate Markdown file under references/. Adding a new platform requires creating one new Markdown file in that directory, not modifying any script. The README states this explicitly: 'one sentence to add a new platform.' The references/ directory currently contains workflows for Xiaohongshu image posts, Xiaohongshu long-form posts, X tweets, and Juejin articles, with the directory listing indicating additional platforms are covered.

Because the workflow is in Markdown rather than code, Claude Code can read and modify it during a session. When a page layout change breaks a step, the agent can diagnose the problem, update the relevant Markdown file, and continue.

Installing social-push and Its Dependencies

social-push requires Claude Code and agent-browser, which in turn requires a Chromium browser. Install agent-browser and its browser dependency first:

npm install -g agent-browser
npx skills add https://github.com/vercel-labs/agent-browser --skill agent-browser
agent-browser install

The agent-browser install command downloads the Chromium testing browser. After that, install the social-push skill:

npx skills add jihe520/social-push

The skill files install into the Claude Code skills directory. The README also mentions a manual copy option: copying the .claude/skills/social-push directory from the repository into your own project if you prefer not to use npx.

To enable remote debugging so agent-browser can connect to the browser, navigate to chrome://inspect/#remote-debugging in a Chrome or Chromium instance and toggle 'Allow remote debugging for this browser instance'.

Platforms, Draft Mode, and First Login

The skill currently supports eight platform targets: Xiaohongshu image posts, Xiaohongshu long-form posts, X tweets, Zhihu thoughts, Weibo posts, WeChat public account articles, Juejin articles, and Linux.do forum posts. The default connection mode is --auto-connect, which attaches to the browser the user is already running rather than a separate Chromium instance. This means the session cookies from the user's regular browser are available.

All posts are created as drafts. The skill explicitly does not publish automatically. After the agent fills in the content fields, the user reviews and confirms before the post goes live. This is a safety design: automated publishing without human review is treated as out of scope.

Platforms that require CAPTCHA-based or slide-verification login cannot be accessed by the automation for the first login. The README provides a manual login flow: launch the Chromium testing browser with remote debugging enabled, connect agent-browser to it, complete the login manually in the visible browser window, then save the session state with agent-browser state save ~/my-state.json.

License and Maintenance Status

No license is declared in the repository. The README has no LICENSE link or terms statement. Default copyright applies, which means using the skill in a commercial product or sharing a modified version requires the author's permission.

The last push was on 2026-06-28. The README mentions the author is building a social media operations agent at gutchain.fun and links to an X account for updates. The skill's maintenance depends on that project's development trajectory. Because each platform workflow is a Markdown file, contributors can update specific platform support without touching the core skill logic.

Editorial conclusion

social-push is a practical choice for a developer who regularly publishes the same content to multiple Chinese and international social platforms and wants Claude Code to handle the browser interaction. The self-evolution design reduces the maintenance burden compared to selector-based scripts. The hard constraints are that it requires agent-browser and its Chromium dependency, the initial login to each platform must be done manually at least once, and the skill creates drafts rather than publishing directly. No license is declared in the repository, so using the skill in a commercial or team context should be cleared with the author first.

Frequently asked questions

Official sources

1. Issues
2. jihe520/social-push on GitHub
3. README