AI video walking prompts: make each step reviewable
Name who walks, where they travel, their pace, and where they finish. Keep the shoes and ground visible when foot contact matters. For an initial trial, use a short, simple walk and a fixed view so you can inspect the steps. A detailed prompt expresses your intended movement; it cannot guarantee correct anatomy or prevent sliding feet.
1. Give the walk a visible route and ending
“Walk naturally” leaves several decisions open. Does the person approach the camera, cross the frame, or walk away? Do they keep moving at the end, or settle into a pause? Choose the result your scene needs before adding a turn, a gesture, or a changing camera angle.
Kling's realistic-video guide recommends one main action per shot and describing important physical interactions, including feet meeting the ground. That is useful planning guidance, rather than a promise that a particular sentence enforces contact in every generation.
A full-body side or three-quarter view can make the legs and surface easier to inspect than a tight portrait crop. Leave room in the intended direction of travel. If a long coat, another person, or foreground furniture hides the shoes, record that you cannot judge the hidden contact from that view. Choose movement appropriate to the character instead of imposing a single walking style on every scene.
A tracking camera can keep a walker in the same part of the picture while the ground moves behind them. That alone is not foot sliding. Review the shoe against the nearby surface during contact, rather than requiring the shoe to stay at one screen coordinate. Plan a camera move after you know what the person should do.
2. Copy a walking-shot card
Use this card before generating, then add observations after watching the whole result. Step counts and pauses are descriptive intentions, not exact timing controls. Check the selected model's actual input mode, duration options and controls separately.
Subject and opening pose: Visible shoes, surface and route: Direction through the frame: Pace and short action: Intended ending or pause: Camera view and allowed movement: Reference file, model/mode and exact prompt: Contact, body travel and identity observations: Accepted / rejected, with the visible reason:
For an original planning scene, imagine an adult in charcoal trousers, a cream shirt and tan lace-up shoes crossing a quiet courtyard. Broad gray paving slabs provide nearby landmarks, and a low stone planter marks the intended stopping area. The landmarks help a person review the take; they do not create a tracking system or lock the generated feet to the ground.
On a small screen, scroll the table sideways.
| Phase | Intended result | What to inspect |
|---|---|---|
| Opening | The person and both shoes are readable above the paving | A usable starting pose and space for the planned route |
| Step | The moving foot advances while the other supports the walk | Whether apparent contact slips, floats or passes through the paving |
| Travel | The body progresses along the intended route | Leg movement that repeats while the body travels inconsistently |
| Ending | The person settles near the planter | A readable pause rather than an unrequested turn or jump in position |
3. Adapt these original walking prompts
These are unrendered draft examples. They have not been benchmarked or shown to outperform another prompt. Use a small trial to ask one clear question, keep its settings, and judge the exported result.
Text-to-video: establish a short courtyard walk
A full-body side view of an adult wearing charcoal trousers, a plain cream shirt and tan lace-up shoes in a quiet courtyard with broad gray paving slabs. The person takes two relaxed steps from left to right toward a low stone planter, then settles into a comfortable pause. Each advancing shoe clears the paving before meeting it again as the body progresses along the route. The camera holds a steady view with the shoes and nearby ground visible.
The subject, surroundings and appearance belong in this version because the text is establishing the scene. The step description supplies an observable intention. It does not activate a gait simulator or prove that the model calculates forces, weight or joint motion correctly.
Image-to-video: describe the motion your source supports
The person takes two relaxed steps toward the open space on the right, then settles into a pause. The body travels with the steps, with clear shoe-to-ground contact. The camera holds its position and keeps the whole person in view.
Use this only when your actual reference shows the intended person, visible footwear and a plausible route. Select that image through the service's supported image input. Naming a file in a prompt does not attach it. A crop that ends at the knees cannot establish the exact appearance of shoes the model later invents.
Runway's Gen-4 image-to-video prompting guide treats the image as visual context and recommends simple descriptions of subject motion. Apply that advice within the documented workflow; other models or modes can expose different controls and requirements.
Simpler alternative: one step and a pause
From the visible standing pose, the person takes one small step forward and settles into a steady pause. The shoes and nearby paving remain visible throughout the fixed full-body view.
This changes the assignment to a shorter action. If it succeeds, it answers whether that simpler movement works in this setup, not whether the original longer walk has been repaired. Keep turns, stairs and other gestures as separate decisions rather than adding them all to the next attempt.
4. Review contact, body travel and appearance separately
- Watch at delivery speed. Does the whole walk read as the intended action, with a coherent route and ending?
- Scrub each visible contact. Compare the shoe and nearby paving as support changes. A lifted foot moving forward is expected; a foot that appears planted but skates along the surface is a different observation.
- Follow the body. Look for a sudden jump, a change of travel direction, or legs cycling without corresponding progress. Account for any intended camera movement.
- Check the character. Inspect shoes, leg shape and clothing independently of the motion. A convincing step can still change a required visual detail.
- Check the ending. Confirm the chosen pause or continuing walk lasts long enough for its intended use.
Write a concrete failure note: “the tan shoe slips along a slab seam during the second apparent contact,” or “the shoes leave the frame before I can assess the last step.” Avoid inferring the model's internal cause from a single clip. Save the input file, exact prompt, settings and output identifier with that observation.
Blur can hide detail without proving that contact is correct. Use the character consistency guide for appearance checks. If you plan slow playback, inspect that delivery version too; the slow-motion guide explains the separate decision between generated movement and retiming accepted footage.
5. Change the next attempt for a specific reason
If the contact is hidden, prepare a reference and framing that reveal it. If travel and camera movement are hard to separate, simplify the camera. If the ending fails after an acceptable opening step, try the shorter action before adding more descriptive words. Keep the other inputs recorded so you know what changed.
Set a trial limit before generating. When precise full-body movement remains essential and attempts keep failing, consider documented motion controls, suitable source footage, or an animation/editing workflow you can actually control. An intentional upper-body crop or a cut to a reaction can serve a different edit, but cannot demonstrate that the hidden steps became correct.
Shotmelo currently organizes manual image and motion briefs, previews a local reference, and saves or exports text drafts. It does not render video, analyze gait, correct feet or retime footage. Copy the walking card into an editable brief, then use your chosen generation or editing service for the actual take.
Published November 2, 2026. Official prompting guidance checked on that date. Card, scene and prompts are original planning examples, with no paid rendering tests or guaranteed motion result.