Miles Ai Video

Miles AI Video

Turn an AI product, workflow, or raw reference into a short video that can actually be shot, edited, rendered, and reviewed. Prefer one true product moment over a polished montage that proves nothing.

Lock the video job

Before writing a shot plan, identify one output mode:

• Demo video: show a real product or workflow result clearly enough that another person understands what happened.
• Reference adaptation: preserve a supplied reference video's information order, rhythm, and proof structure while replacing only the case material.
• Script-to-edit plan: convert a spoken script or rough outline into scenes, captures, overlays, and edit checks.
• Review: inspect an existing cut and name exact seconds where meaning, pacing, legibility, or proof breaks.

Use only supplied or verified material. Do not invent product results, usage numbers, platform performance, screenshots, quotes, dates, revenue, or publication status. If the strongest claim is not proven, narrow it.

Find the one proof moment

State the video promise in one sentence: after watching, what can the viewer now believe or try?

Then choose one proof moment:

• the before state the viewer recognizes;
• the action the maker performs;
• the output that visibly changes the state;
• the boundary of what the video does not prove.

Delete clips, effects, transitions, and lines that do not serve that proof. A video can be pretty and still fail if the viewer cannot tell what changed.

Build the edit spine

Produce a compact plan with these parts:

1. Opening conflict: the exact pain or curiosity in the first 3 seconds.
2. Screen proof: what screen, file, prompt, or product state must appear.
3. Human action: what the maker says, taps, drags, types, or records.
4. Result reveal: what the viewer sees after the operation.
5. Boundary close: what this demo is not claiming.

Keep shots practical. Prefer phone screen recordings, desktop captures, one face-to-camera bridge, and readable overlays. Do not require studio gear, custom animation, expensive stock footage, or private account access unless the user explicitly provided it.

Output the usable artifact

Default deliverable:

• a 30-90 second structure;
• shot list with approximate seconds;
• on-screen text for each segment;
• voiceover or talking points;
• required captures/assets;
• acceptance checks for legibility, timing, and evidence.

For a review, return exact timestamps, the problem at each timestamp, and the smallest edit that fixes it.

Quality gates

Block delivery for hard errors:

• the claimed result is not visible;
• a screenshot or clip shows the wrong product;
• private data, secrets, chats, or unapproved faces appear;
• text is unreadable on mobile;
• the story implies revenue, growth, user adoption, or publication proof that was not verified;
• the video asks the viewer to trust narration where a screen result should be shown.

Treat wording, music, color, and local pacing as soft problems. Fix them locally without throwing away a working proof structure.

Protect Miles personal information

Keep Miles Wang, Add-Miles, and the public repository identity as allowed brand references. Before returning or exporting anything, remove non-brand Miles contact details, private paths, account identifiers, chats, unapproved faces or voices, and unnecessary source-media metadata. Stop if the task would require exposing them.

Respect the router boundary

When mileswang invokes this Skill, use the host-provided active Skill catalog as the only availability authority and keep the selected executor unchanged. Do not rediscover or replace the selected executor from disk folders, plugin caches, local inventories, or configuration files.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Add-Miles
• Source: Add-Miles/mileswang-skill
• License: MIT

Install and usage instructions live in the source repository linked above.