I built PromptClip-Skill, an open-source Codex Skill for finding the meaningful moments inside a folder of casual videos.
This started with a very ordinary problem: recording family life is easy, but reviewing it is not.
A few minutes of a baby playing, a beach walk, or a weekend outing can produce dozens of clips. Many contain camera shake, accidental recordings, repeated actions, empty frames, or a subject leaving the frame. The moments worth sharing are there, but buried in hours of low-signal footage.
The key idea
I think of this as a filtering problem before it is an editing problem.
Instead of asking an AI editor to create a generic montage from every file, I give it a natural-language selection rule:
Keep moments where the baby is clearly visible, reacting naturally, laughing, interacting with a parent, or completing a meaningful action. Remove shaky footage, repeated scenes, empty frames, and incomplete actions. Preserve the original order.
The prompt describes both positive signals and exclusions. The output is a shortlist of candidate clips with timestamps and an edit decision list that can be reviewed before export.
That makes the workflow useful for personal memories, where selection quality and transparency matter more than automatic transitions.
Why I built this
One-click video creation tools are convenient, but they can be difficult to control for this specific use case:
- short preset outputs can cut away too much context
- custom prompts are not always followed precisely
- negative requirements such as "remove all repeated or empty footage" are hard to express
- it is not always clear why a clip was selected
- the result often follows a generic template instead of the structure of the original memory
For a beach walk with a child, I may want the first discovery of the water, a clear facial expression, a complete interaction with a parent, and a few coherent walking shots. I do not necessarily want every visually active second.
Workflow
PromptClip-Skill runs locally in Codex:
- Point it at a local folder of videos.
- Describe the people, actions, mood, duration, and exclusions.
- Analyze the footage and identify candidate moments.
- Review timestamps and the generated edit decision list.
- Export selected clips with FFmpeg or continue editing in another tool.
The skill does not modify the original media. Intermediate results remain inspectable, so the selection can be adjusted when the prompt needs refinement.
It is not tied to family videos. The same approach can be used for travel, pets, sports, events, interviews, lessons, or any collection where useful moments are surrounded by throwaway footage.
Project status
This is a small open-source experiment, not a hosted video SaaS or a replacement for a professional NLE. It currently focuses on making prompt-based selection practical and repeatable.
Repository: https://github.com/ron0115/PromptClip-Skill
I would especially value feedback on better highlight scoring, prompt templates, local model support, and ways to make the review/export step more reliable.
The larger question I am exploring is simple: can an AI editor be more useful when it first learns what to throw away?
Top comments (0)