TikTok automation

This project is reuse-first. Before adding a new workaround, use existing mature capabilities: pipeline modules, provider adapters, download-collector, retry commands, dreamina CLI, lark-cli, enabled Codex plugins, and installed skills.

Browser defaults now use a shared OpenClaw-style persistent Chrome family: TikTok monitoring runs in a headless automation-owned profile, while ChatGPT image generation/review/download runs in a separate headed automation-owned profile. Both clone from the same shared source profile so they can stay logged in at the same time without sharing a live run profile.

Browser operations use the shared persistent-browser-split-runtime-v1 policy: OpenClaw-style profile clone first, short DOM/media reads first, screenshots only as a fallback or handoff aid, and separate headed/headless run profiles so ChatGPT and TikTok tasks can run concurrently.

For ChatGPT web images, review the generated images in the page before downloading. Download only accepted images into the package; retry mismatched shots in the same conversation.

Generated assets are reviewed for TikTok 钩子 strength as well as prompt match. Failures must produce reusable lessons in review logs or prompt presets so future scripts reuse what was learned. For business 钩子, a calm accurate scene can still fail if the reference-level lure is stronger. Opening beats should usually push visible money shock, cash rain, status contrast, or stunned reactions before quieter explanatory scenes. If ChatGPT returns multiple independent images displayed as a grid of thumbnails, review them individually as batch output. If it returns one image that contains a multi-panel grid/storyboard, treat it as preview-only and regenerate accepted cells as standalone 9:16 images before Dreamina image-to-video. Image prompts should not reserve blank subtitle space. Captions are added over the image in CapCut, so assets should stay full-frame and visually complete.

Built With

• html
• javascript
• powershell