Reverse-engineer prompts from AI videos you admire

What it does

PromptLens is a web dashboard that ingests videos or images and uses vision-capable LLMs to reverse-engineer text prompts describing the scenes, styles, and actions. It also transcribes audio to extract dialogue, and can clip segments via a self-hosted FFmpeg service. Think of it as a prompt extraction and remixing factory for the AI video era.

The interesting bit

The project openly embraces the “copy to learn” workflow: it doesn’t just analyze, it structures the output as reusable prompts you can feed straight into video generators like Kie.ai. The roadmap even includes automatic shot detection using FFmpeg scene thresholds and PySceneDetect to chop long videos into analyzable chunks.

Key highlights

• Supports multiple vision LLM backends (Zhipu AI, Gemini, OpenRouter) so you aren’t locked into one provider.
• Audio transcription offers both a managed cloud option and a fully self-hosted, free alternative via FunASR.
• Built as a modern Next.js 15 stack with PostgreSQL, Drizzle ORM, and Cloudflare R2 for storage.
• Includes user auth, history tracking, and operation logging for a multi-tenant workflow.
• The video editing feature requires deploying a separate HTTP-wrapped FFmpeg service; the README candidly suggests casual users just use CapCut instead.

Caveats

• Video editing is not turnkey: you must self-host an FFmpeg HTTP service, and the authors admit Vercel can’t handle it.
• Long videos yield vague prompts; the tool practically demands you pre-cut footage into 6–10 second clips.
• The built-in OpenRouter key is convenient but shared, so expect rate limits unless you bring your own API credentials.

Verdict

Content creators and AI video tinkerers who want to dissect and remix existing clips will find this a useful starting point. If you need a fully managed, zero-config video editor, this isn’t it—it’s a developer-oriented glue layer that expects you to supply your own API keys and FFmpeg worker.