AI Mendao · Lingke
A lightweight AI director for video creators, exploring a shorter path to one-click video production.
Workflow diagram reconstructed from the project account.
Fewer handoffs between a story and a finished video
AI Mendao · Lingke explored a lightweight AI director for video creators and a shorter path to one-click production. Story development, shot planning, generation and editing often sit in different tools, with information repeatedly copied between them. The product aimed to reduce that work so creators could spend more attention on what they wanted to express.
I began making AIGC videos manually in September 2025, then developed the product from October 2025 to March 2026 as founder and product lead. It is now offline. This case records workflows and tradeoffs we tried. One-click production was the direction explored; complete, fully automated production had not been validated.
The director Agent needs to explain each shot
The workflow began with a story. A director Agent organized shots, the timeline and shot prompts, then called the Sora 2 service available to us at the time to generate clips. FFmpeg assembled them. That service had a 15-second limit per shot, so longer material needed several shots.
The difficult part was deciding what each stage needed to hand over. A story saying “the atmosphere is tense” still leaves questions: who is present, what happens, and what change should this shot show? Making the prompt longer may leave those questions unanswered.
I mainly looked at whether the director Agent's plan could be executed: what each shot should convey, whether its duration made sense, and whether adjacent shots had shared references. A beautiful clip could still fail to fit the story, leaving the editing stage to deal with it.
Why does the next shot have to wait?
We used a serial method that took the final frame of one shot as the first frame of the next. This gave the next shot a visual reference for continuity, but required it to wait until the preceding clip was finished.
Users wait for the complete video. If shots keep waiting for each other, faster individual generations may still leave a long wait for the final result. Some of that delay came from the model; some came from how we arranged the work.
Revisions had the same dependency. Changing an earlier shot could affect later shots that used it as a reference. Beyond the normal generation order, we needed to work out where to resume when a step went wrong.
Prepare references first, then generate the clips separately
The later workflow derived the full screenplay and first-frame image prompts for all shots, generated reference images, and then produced video clips in parallel. This workflow ran successfully in n8n.
With characters, settings and shot references prepared upfront, the clips could begin separately instead of each waiting for the previous clip's final frame. Changing those dependencies removed one source of waiting.
We still needed to check whether the reference images agreed and whether characters and settings remained consistent. Parallel work did not solve every continuity problem. There is no confirmed quantitative speedup to report, and a working n8n flow does not mean it was fully integrated into the production product.
Fewer buttons should still leave a way to recover
To me, lightweight design makes it easier for creators to understand what they are doing, how far the system has progressed, and what still needs their judgment. Leaving one button on screen is not much help if an error leaves them guessing what happened.
The product still needs to explain some intermediate results: how the story was divided into shots, which materials are ready, where something failed, and what a revision affects. We can show fewer technical details without hiding the information someone needs to continue.
When improving a video workflow again, I would first look at task dependencies and whether failures can be inspected and recovered from. Then I would consider how to speed it up and which operations can be removed. The eventual measure is the whole work: its quality, how long it took, and how much had to be redone.