Prompt Ayesha
AI Video Prompts•3 min read

The Hand That Never Touches the Packet

Prompt Ayesha•
The Hand That Never Touches the Packet

You upload two character images, paste the prompt, and hit generate. The first pass looks promising: the right person is eating chips, the left person is sulking. Then you watch the sneaky reach. The hand goes in. A chip disappears. The entire joke collapses because the one thing that must never happen—the theft—just happened.

This is not a prompt-clarity problem. Your prompt says “EMPTY” three times and “NEVER” in all caps. Google Flow’s video model reads those words, understands them, and still produces a hand entering a packet. The failure lives somewhere deeper: in how the model interprets “reaching toward” as a continuous action with a completion state.

The hard problem here is action negation in a reference-driven video. You are asking the model to preserve a motion trajectory—the reach—while forbidding its natural endpoint. Most video models treat “reach toward X” and “take X” as a single learned sequence. Stopping mid-sequence requires a different kind of instruction.

Fix 1: Separate the Reach from the Grip

The model fails because your prompt describes a reach and a non-take in the same beat. When the text says “slowly reaches his hand toward the open packet,” the model’s training data overwhelmingly associates that motion with contact. The negation (“never takes a chip”) arrives after the action description, so the action wins.

Restructure the instruction so the hand’s endpoint is defined as a spatial coordinate, not a behavioral negation. Instead of “reaches toward the packet but does not take,” write: “his hand stops in the air six inches from the packet opening, fingers extended but not flexed, then pulls back.” The model can render a fixed distance. It struggles with “don’t complete this motion.”

In Flow, reference images anchor identity, not action logic . The prompt must carry the full burden of where the hand stops. Give the model a visual target—air, not packet—and the empty-hand rule becomes easier to enforce.

Fix 2: Lock Identity with a Single Reference Per Character

Your prompt asks Flow to preserve “exact facial identity, facial features, hairstyle, skin tone, clothing, and appearance” for both characters. That is six identity dimensions per person, twelve total, across six story beats. Flow’s reference system works best when each character has one clean, frontal reference image on a plain background . If your uploaded images have complex lighting, partial occlusion, or different angles, the model interpolates—and interpolation means drift.

Use one reference per character. Not two. Not a collage. The left-sitting character gets one image, the right-sitting character gets another. Name them explicitly in the prompt with Flow’s @ syntax: @1image and @2image work, but Flow also accepts character names if you have set them up in the project . The model tracks named entities more reliably than positional references like “first character.”

If identity still drifts between the reach beat and the catch beat, the problem is beat length. Flow’s video model handles 4–6 second generations more stably than longer ones for action-heavy sequences . Split the six-beat story into two generations: beats 1–3, then beats 4–6. Use the last frame of the first generation as the starting frame for the second. This gives the model a visual anchor for where the hand was, rather than asking it to remember across a long timeline.

Fix 3: Make the Catcher’s Face the Action Trigger

Beat 4 is the pivot: “the second character turns and catches him.” In your current prompt, this is described as a narrative event. The model has to infer when the turn happens, how fast, and what expression accompanies it. That inference is where timing drifts.

Rewrite the beat as a facial action that triggers the hand retraction. “The right character’s head snaps toward the left character, eyes narrow, mouth stops chewing. At the exact frame her eyes land on his hand, his fingers curl inward and pull back to his chest.” The model now has a causal chain: her gaze movement causes his withdrawal. This is easier to render than two independent actions happening at approximately the same time.

The suspicious stare in beat 5 and the arm-around-shoulder in beat 6 can be chained to the same logic. Her stare does not relax. His arm moves because her stare has not relaxed. Every action in the sequence has a visible cause in the other character’s face or body.

Fix 4: Anchor the Packet as a Fixed Object with a Closed State

The green chips packet is not a character, but it behaves like one in this prompt. It must stay in the right character’s hand, remain open, and never receive the left character’s fingers. Flow’s reference system supports object references, but the packet is already in the video you are uploading as a motion reference .

The problem is that the model may treat the packet as a receptacle that invites interaction. To counter this, describe the packet’s state at every beat where the left hand moves. “The green packet stays tilted toward the right character’s chest, opening facing away from the left character’s reach.” This spatial relationship—opening direction—gives the model a physical reason the hand cannot enter, even if the reach continues.

During the sneaky reach beat, add: “The packet does not move toward his hand. It stays in her grip.” When the model has to choose between moving the packet to meet the hand and stopping the hand short, the explicit “does not move” instruction gives it permission to leave the hand empty.

The Prompt

Recreate the uploaded reference video exactly. Replace the left-sitting character with @1image and the right-sitting character with @2image. Preserve their exact faces, hair, skin tone, and clothing from the reference images. CRITICAL CONSTRAINT: The left character’s hand NEVER enters the green chips packet. He NEVER holds, takes, or eats a chip. His hand stays completely empty for the entire video. BEAT 1: Right character eats chips from the green packet, laughing. Left character sits with chin on hand, watching the packet. BEAT 2: Right character tilts the packet toward him. Left character shakes his head once, small motion, and looks away. BEAT 3: Right character looks forward, chewing. Left character slowly extends his right hand toward the packet. His fingers stop six inches from the opening, hovering in air. BEAT 4: Right character’s head snaps toward him. Her eyes narrow. At the moment her gaze lands on his hovering hand, his fingers curl inward and pull back to his chest. The packet does not move. BEAT 5: Left character raises his empty hand, palm open, and makes small circular gestures as if explaining. Right character stares at him, mouth closed, eyes flat. BEAT 6: Left character leans toward her and places his arm around her shoulder. His other hand remains visible and empty. She does not smile. Same background, same camera angles, same lighting, same timing as the reference video. No watermarks. No text overlays.

The prompt works because every “never” is paired with a positive spatial instruction. The hand does not just avoid the packet—it stops at six inches, then curls, then pulls to the chest, then raises open-palmed, then rests on a shoulder. The model has five distinct hand states to render, none of which involve gripping. The packet’s immobility is stated at the moment of temptation, not buried in a general rule at the top.

Running It

Use Gemini Omni Flash or Veo 3.1 in Flow for this sequence. Both support video-to-video editing with reference images . Upload your reference video as the motion source and the two character images as ingredients. Do not exceed 10 seconds per generation if you are editing an existing clip—Flow’s editing window caps at 10 seconds for most tiers . For a six-beat sequence, plan two generations: one for beats 1–3, one for beats 4–6, with frame bridging between them.

Check the left character’s hand first. Pause on the frame where the reach happens. If the fingers are inside the packet boundary, the generation failed regardless of how good the faces look. Check the right character’s eyes second—if she is smiling during the catch, the suspicious tone is gone and the ending will not land.

Expect three to five attempts before the empty-hand rule holds across all beats. Identity drift between attempts is normal; the model re-rolls facial details with each generation. Pick the attempt where the hand behavior is correct, then use that output as the new motion reference for a refinement pass focused only on faces.

When It Goes Wrong

Hand enters packet despite “six inches” instruction. → The model is interpreting “toward” as “into.” Change to “his hand rises to shoulder height and stops. It does not descend.” Remove the word “packet” from the reach sentence entirely.

Right character offers chips with a smile, left character shakes head, but the packet disappears in the next beat. → Object persistence failure. Add “The green packet remains in her hand at all times” to the end of every beat description.

Left character’s face changes between beat 3 and beat 4. → Identity drift from action complexity. Split the generation at beat 3/4 and use the final frame of the first clip as the start frame of the second .

Right character’s suspicious stare looks playful instead of annoyed. → Expression vagueness. Replace “serious, suspicious, annoyed look” with “mouth closed, eyebrows flat, eyes not blinking, head tilted five degrees down.”

Arm-around-shoulder beat produces a hug or a pull. → Motion overshoot. Specify “his hand rests on her far shoulder, fingers relaxed, no pulling motion. Her body does not move toward him.”

The entire video has a pasted-on look where the new faces do not match the lighting of the original. → Reference image mismatch. Use character images with lighting that matches the reference video’s direction. If the video has side light from the left, the reference image should not be front-lit.

First character’s raised hand in beat 5 holds an invisible chip. → The model hallucinated an object because the gesture context implies one. Add “his fingers are open and empty. No object is between them. The palm faces the camera.”

Timing is compressed—beats 1–3 happen in two seconds. → Flow’s default generation speed does not match your reference. Specify durations: “Beat 1 lasts two seconds. Beat 2 lasts one second. Beat 3 lasts three seconds.” Or use the reference video’s timestamps if you can measure them.

Output has a watermark or logo from the reference video. → The model copied it. Add “No watermarks, no logos, no text anywhere in the frame. The output must be clean.” Place this at the end of the prompt, not the beginning.

Beyond the Chips Packet

The empty-hand problem generalizes to any prompt where a character must approach an object without completing the interaction. A child reaching for a cookie jar but stopping when a parent enters. A thief’s hand hovering over a wallet. A person about to knock on a door but hesitating.

The technique is the same: replace behavioral negation (“does not take”) with spatial coordinates (“stops at X distance”), chain actions to visible triggers in another character’s face or body, and split long sequences at the moment of highest action complexity. Flow’s reference system handles identity well enough when each character has one clean image . The model’s weakness is not faces. It is the space between intention and completion, where the motion lives but the outcome must not.

Give it a fixed endpoint in air. The hand will stay empty.

Share this creative guide:
PA

Prompt Ayesha Editorial Desk

Curated daily AI photography prompts, Lightroom presets, and video editing transitions engineered to help creators elevate their visual content across Midjourney, Bing Copilot, and Flux.

You Might Also Like

More in AI Video Prompts →
Navratri Viral Video Prompt for Flow AI
AI Video Prompts
•Prompt Ayesha

Four Outfits, Ten Seconds, One Fixed Camera — The Multi-Outfit Video Problem

Video prompts on this site usually solve a motion problem — how does the camera move, how does the person move, how do they interact with the scene. This one is the opposite. The camera doesn’t move. The person barely moves. Everything else — outfit, jewelry, styling, mood — changes four times in 10 seconds. […]

Realistic Mercedes AI video prompt featuring a man inside a black luxury car
AI Video Prompts
•Prompt Ayesha

Reference Video Beats Reference Image — Prompting for AI Video That Looks Real

Every other post on this site is about generating a single image from a prompt and a reference photo. This one is different. Video prompts work on a completely different principle — you’re not just describing what should be in the frame, you’re describing what should happen and when. That’s the shift most people miss. […]