I Trained Claude Code To Run My X Account (no API)

A hands-on workflow shows how Claude Code can be trained to run an X account autonomously—without using an API—by iteratively “learning” repeatable browser actions and saving them as reusable skills. The demonstration starts with a logged-in X session where Claude Code searches for recent posts about “Claude Code,” analyzes what’s trending, and then drafts and posts a meme. It navigates from search results to composing, generates an image file (e.g., “claude meme.png”), uploads it, adds text, and submits the post—while continuously writing commands in the background. The key takeaway isn’t just that automation works; it’s that the automation is built through a training loop: set a goal, attempt the workflow in a real browser, retry when needed, and then codify the successful steps into a skills file so the agent can repeat the same behavior later.

From there, the training expands into a second capability: video understanding on X. The agent is tasked with handling two different cases—videos without audio and videos with audio—because X content varies widely. For an audio-less clip, Claude Code first downloads the video from X using a YTDLP-based approach, then probes the file with FFmpeg to confirm whether audio exists. When audio is missing, the workflow shifts to extracting frames at a fixed interval (for example, one frame every few seconds) and using those visual samples as context to summarize what the video is about. In the example shown, the frame-based analysis produces a coherent description of a “snake game” demo.

For videos that include audio, the workflow becomes more text-centric. Claude Code again downloads the video, converts it to MP3 using FFmpeg, and then extracts speech using a Whisper-based transcription step. The resulting audio transcript is saved as a text file (while the MP3 is removed afterward), and that transcript becomes the agent’s context for downstream tasks like summarization, research, or generating comments. To make the output easier to consume, Claude Code also generates a simple index.html page that presents the video’s takeaways in a browser-friendly format.

The broader system design is a skills library: once a workflow reliably works, the steps are written into a skills.md file so the agent can reuse them in future runs. The transcript also mentions running Claude Code in a “dangerously skip permissions” mode and coordinating multiple versions in parallel, emphasizing speed and autonomy. Finally, there’s a playful but cautionary side project: a “skills store” (skills.md.store) where trained skills can be downloaded after payment, framed as a meme for now but flagged as potentially risky if users download files from untrusted sources.

Overall, the core insight is practical: robust X automation is built by training browser-level tasks into reusable skills, then layering specialized pipelines—like frame extraction for silent videos and Whisper transcription for audio videos—so the agent can interpret content and act on it consistently.