- Home
- Skills
- Image & Video
- AI Video Shot Check: Catch Fake Physics Before You Render
More screenshots
Works with the AI Agents you already use
AI Video Shot Check: Catch Fake Physics Before You Render
Check your AI video shot list and image-to-video prompts (Kling, Veo, Runway) for fake physics before you render; get a fixed prompt per failed shot.
$6
AI Video Shot Check: Catch Fake Physics Before You Render
Example session with this skill installed
check these before I render:
Shot 1 – START: closed kitchen drawer, close-up. END: drawer open, spoons visible. Motion: "Camera pushes in slowly as the drawer slides open revealing the spoons."
Shot 2 – START: a corgi sitting by a low coffee table, its tail resting against the table edge, a folded napkin on the edge right above it. END: napkin on the floor. Motion: "The corgi wags its tail, which sweeps the napkin off the table; the napkin flutters to the floor. Camera locked."
Shot 3 – START: takeaway cup with a sleeve printed "HOUSE BLEND". Motion: "Slow orbit around the cup as the camera tracks across the sleeve, steam rising."
- Read your context and instructions
- Compiled the ai video shot
3 shots: 1 PASS, 2 FAIL · Checked: text-only (stills not inspected) · critic: separate subagent
| Shot | Verdict | Rules | Fix in a few words |
|---|---|---|---|
| 1 | FAIL | P1 | a hand pulls the drawer open |
| 2 | PASS | – | – |
| 3 | FAIL | P6 | camera locked, only steam moves |
Shot 1 – FAIL (P1, nothing acts by itself)
What a viewer would see: a drawer opening on its own.
Evidence: "the drawer slides open"
Fix: a hand enters, pulls the handle, lets go.
Camera locked. A hand enters from the bottom edge, takes the drawer handle and pulls the drawer open, revealing the spoons, then lets go and moves out of frame. The counter and the spoons stay still.
Shot 2 – PASS. The napkin moves, but the tail sweep that moves it is in the prompt and START puts the tail under it. Critic flagged a jump between START and END; overruled, the cause is stated.
Shot 3 – FAIL (P6, text stays still)
What a viewer would see: "HOUSE BLEND" smearing as the camera circles the cup.
Evidence: "Slow orbit around the cup as the camera tracks across the sleeve"
Fix: lock the camera with the sleeve facing it; only the steam moves. Keep the text sharp in the START still.
Camera locked, no zoom. Steam rises from the cup and drifts slowly to the left. The cup and its printed sleeve stay still, facing the camera.
Not checked: no images attached, so the stills themselves weren't inspected.
Connects securely to your tools. The creator never sees your data.
About this skill
AI Video Shot Check reads your shot plan before you spend credits on renders. Paste a shot list, the descriptions of your START and END frames, or the motion prompts you're about to give an image-to-video tool (Kling, Veo, Runway, Seedance or any other).
For every shot it asks one question: could a real camera have filmed this, starting from this exact frame? It catches the usual failures: a box that opens with nobody touching it, a cup that sits in one place in the first frame and somewhere else in the last, a hand "placing" a photo that is already on the table, "turns into" transformations, a thumb sinking into a peach, a label smearing as the camera circles it, objects floating without support, and the hero product changing look between shots.
Motion with a visible cause passes. A cat knocking the yarn off the sofa is physics, and the skill is built to avoid false alarms as much as misses.
Each shot gets PASS, PASS with a risk note, or FAIL. Every FAIL names the rule, quotes the words that cause it, gives the fix and a rewritten motion prompt you can paste straight into your tool, plus a still change when the prompt alone can't fix it. An independent critic checks the same shots blind, and every rewrite is checked again before you see it.
Optional: notes on your still prompts (short prose, one named light, reference images for anything that must stay identical, exact text in quotes).
Text-only works fully on descriptions and prompts. If your agent can see images, it also compares the START and END stills.
It checks plausibility before you render. The check itself doesn't render or spend anything. It can't make a model obey a prompt. If you want a clip afterwards, your agent can use the Agensi video tools listed on this page; each one shows its price and asks first. The check uses no scripts and no network access; your optional settings live in ~/.ai-video-shot-check/config.md or in the chat.
- It checks the plan before rendering, so it can't see how a specific model will interpret a verb, the exact timing or motion blur, and a PASS doesn't guarantee the model follows the prompt. Without image access it checks descriptions and prompts only and can't spot problems that exist only in the pixels (a sixth finger, garbled text in the still). The rules are working practice from image-to-video production; where your tool's own documentation differs, follow the documentation. It judges physics and continuity only; story, taste and brand stay with you. Without subagents the independent critic runs in the same chat and is less independent. The check itself doesn't render or spend anything, so it can't show you the motion. If you want a clip afterwards, your agent can use the Agensi video tools listed on this page; each one shows its price and asks first. The skill can't vouch for how a clip turns out.
- Nine shot rules (nothing acts by itself, no unmotivated jumps, motion continues the start frame, one action without transformations, contact with soft objects lives in the still, text stays still, hands, gravity and scale, achievable from this frame) and three sequence rules (same object same look, matching cuts, END made from START).
- Output: a summary table, a card per failed shot with a rewritten prompt in a code block, sequence notes, optional still prompt notes and a list of what wasn't checked.
- The skill has SKILL.md, the rules, motion and still prompt guides, the critic prompt, a config field reference and a fictional worked example; your agent loads it all from Agensi as text.
- Single-user license.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 7 days ago
Permissions
File Scopes
Reads what you paste or attach: shot lists, START and END descriptions, motion and still prompts, and still images or frame grabs if you give them (attached, or a file path in Claude Code). Reads your optional settings from ~/.ai-video-shot-check/config.md, a folder you choose, or the chat. Writes only your own files: config.md if you ask for setup, and a Markdown report only if you ask for one, saved where you say and never over an earlier report. Nothing from the skill itself is ever written to disk. The check uses no scripts, no network access and no connection to any image or video tool. If you want a clip afterwards, your agent can use the Agensi video tools listed on this page; each one shows its price and asks first. Those calls go through your agent's Agensi connection. It never asks for logins or API keys, and the check never asks for credits.