Jake Moran’s AI Video Workflow Starts With a Story and Good Timestamps

RESUMEN

En 30 segundos

Jake Moran uses AI to edit video, but says the creator still needs a rough idea of the story. He recommends precise transcription timestamps for cutting speech, a storyboard before rendering motion graphics, and examples and brand assets to guide the look. In a launch-video demonstration he starts with three screenshots; in a talking-head edit he removes pauses and retakes before adding a restrained introduction. Peter Yang also describes blurring sensitive information and splitting long videos into sections.

1A rough cut removes mistakes before adding effectsMoran asks an agent to remove filler, pauses, repetition and retakes. He says specific language and examples help the model make the intended cuts.
2Accurate transcription timing matters for clean editsMoran uses the local Parakeet model for timestamps. He says inaccurate timing can cut off part of a word, particularly in fast-paced short videos.
3Three screenshots and a story started a launch videoMoran supplied screenshots and a rough narrative, then reviewed a storyboard before the full render. He says the creator still needs an idea of what the video should convey.
4Let the speaker do the work in a talking-head videoMoran used reference screenshots and brand colors for Yang’s introduction. Both prefer effects that support the speaker rather than overwhelm the footage.
5A reference clip gives an agent more context than a description aloneMoran demonstrates adapting a short reference with new type and colors. He says an existing creator can similarly use their own past work as a reference.
6Mark feedback on the frame instead of describing every timestampMoran shows drawing on a frame and queuing notes. Yang adds a blur pass for sensitive information and edits long recordings in sections.

Conclusiones

Describe the story and provide assets and references before asking an agent to edit.
Use accurate word timestamps for speech cuts and review the result for clipped words.
Review a storyboard before rendering an entire motion-graphics video.
Specify brand colors, fonts, desired pacing and music volume rather than assuming the agent knows them.
For long videos, Yang edits sections separately and joins them; Moran saves repeatable workflows for reuse.

TemasGenerative media Visión y vídeo Workflow automation

TEXTO COMPLETO

Lee el texto editado por capítulos

Ver transcripción

Solo la transcripción

From raw recording to rough cut

Moran explains how AI can remove filler, pauses and retakes. Accurate transcription timing is essential for clean cuts.

Peter Yang welcomes Jake Moran of HyperFrames to discuss editing product launches, talking-head videos and short clips with AI. Moran says video creation covers many formats; a raw recording generally runs longer than the finished edit because it includes mistakes and pauses.

Moran uses HyperFrames, an HTML-based editor that agents can work with. He starts by asking for a rough cut: remove mistakes and return to a consistent storyline before adding visual effects.

A transcript with timestamps can help identify filler words, pauses, repeated phrases and retakes. Moran suggests naming what should be removed and giving examples, such as cutting “ums” or shortening pauses to around a quarter of a second. More context helps the agent approach the intended result.

Cutting speech can produce a visible jump when the speaker’s position changes between the two retained pieces. The speakers also discuss using footage from elsewhere in a recording to add context at the appropriate moment without recording again.

Moran uses the locally run Parakeet transcription model because accurate timestamps matter for speech editing. If timings are wrong, the model may cut off part of a word even when asked for the right pause length. He says this is especially noticeable in fast-paced short videos.

Building an introduction with footage and sound

The speakers discuss zooms, B-roll, transitions and music. Moran recommends specifying the desired energy and audio levels.

For an introduction or hook, Moran suggests choosing an engaging moment and adding visual treatment if it supports the story. A gradual push-in, pull-back or cut to a closer framing can direct attention. The speed and easing of the movement affect how it feels.

B-roll can show what the speaker is discussing, establish context or preview a tutorial’s contents. A simple cut moves directly between shots, while a whip pan creates the appearance of a quick camera movement between them. Yang suggests naming effects he likes and allowing the agent to choose where to use them.

Moran says an agent can use supplied music or create music with appropriate tools. He finds that models may place background music too loudly, so he gives an explicit level, such as a fraction of the main speech volume. Giving the agent these constraints up front can save slow editing iterations.

A launch video from three screenshots

Moran walks through a launch video, reviewing a storyboard before rendering. He explains how avatars and brand assets supplied context.

Moran plays a launch video for HyperFrames’ background-removal feature and then explains how he made it. He began with three screenshots, including an image of the studio and the assets for the opening, plus a prompt describing the intended story.

He says creators still need an idea of what they want to communicate, even if they no longer specify every cut. An engineer’s demonstration of background removal gave him the opening. He supplied that asset and described the message; the model made many of the detailed choices.

Before rendering a full motion-graphics video, Moran prefers a storyboard of still frames. It allows him to identify problems and revise the visual direction quickly. The storyboard need not be perfect, but it should be close enough to justify the longer build.

For a feature launch, Moran typically aims for roughly thirty seconds to a little over a minute. He describes four broad parts: an engaging hook, context showing the feature, the payoff or action, and a return to the brand. The structure is a guide to the story rather than a fixed list of effects.

For the example, Moran used HeyGen-generated avatars rather than sourcing real footage for every person in a grid. His own appearance was an avatar too, and he says the model included it without an explicit request. He notes that creators can instead provide real footage.

The HyperFrames desktop app connects to generative models; outside that environment, Moran says an agent needs an appropriate connector or command-line integration. For a consistent look, he provides a brand kit from a website, document or folder rather than assuming the model knows the brand.

Editing a talking-head video and adapting references

Moran cuts Yang’s raw recording and builds an introduction from references. He shows how a short reference clip can guide a new style.

Moran also edited one of Yang’s raw videos. The recording was 20 minutes 30 seconds and the finished edit was 18 minutes 11 seconds after pauses, retakes and other unnecessary material were removed. The amount cut depends on the original recording.

For its introduction he wanted a short reel that showed the tutorial’s use cases. Yang had supplied two visual references, and Moran gave the agent screenshots and brand colors. He describes the desired result rather than manually choosing every cut.

Yang says motion graphics should not overwhelm the person speaking. Moran agrees that a small amount can go a long way: for a talking-head tutorial, the main footage should carry the explanation, even though some creators prefer a busier style.

Short social videos may use more effects to compete with scrolling, while long YouTube videos often establish the viewer’s commitment through the title and thumbnail. Moran says a link, available assets and a description of the desired energy—such as fast-paced or restrained—can help guide an agent.

To reproduce a visual style, Moran previously watched references slowly and described details such as text entrances and background removal. He now demonstrates uploading a short reference clip and asking the agent to adapt it with different colors, type and content.

A reference can be a screenshot, a prior video or a project containing reusable code. Moran says experienced creators can use their own existing library to carry established visual choices into AI-assisted editing. The agent can also use generated or sourced supplementary footage.

Reusable examples and visual feedback

Moran demonstrates reusable launch-video sections and frame annotations. Yang describes privacy and long-video workflows.

For launch videos, HyperFrames provides reusable examples of sections that Moran calls “moments.” A creator can select examples and describe whether to adapt them to the brand, use several for a new story or borrow some of their energy. He says seeing the examples makes them easier to use than finding hidden code in a repository.

Yang asks how to give feedback when a first draft is not right. Moran demonstrates drawing directly on a frame and adding a note; the notes queue until the viewer is ready to send them. He finds this easier than watching while typing timestamps into a chat.

Yang runs a blur pass over screen recordings to conceal sensitive information. For long videos, he splits the work into sections, edits them separately and then joins them, avoiding repeated processing of an entire recording for a local change.

Moran recommends saving a repeatable workflow as a skill or reference after figuring out the steps. This can prevent a creator with a consistent format from repeating the same setup. Yang says AI editing broadens access even though professional editors may still do a better job.

Moran says HyperFrames Studio is available for Mac and Linux at the time of the interview, with Windows support in progress. He also points to the project repository for people who want to use it with an agent.