ai-video-generation

AI Video Generation

A good AI clip is decided before anything renders: one subject, one action, one camera move, in one short beat. This skill writes that shot, prices it, gets a yes, and has the Wireflow app skill-ai-video-generation render it with MiniMax Hailuo 03 Max Turbo.

What the app takes and returns

The format is fixed: 16:9 and 5 seconds. If the person needs vertical 9:16 or a longer clip, say so before you plan, and offer the board link in the report step, where they can change those settings.

1. Check the connection

The Wireflow tools end in run_app, get_execution, get_credit_balance and whoami (the prefix varies with the client). When they are not there:

1. Tell the person generation needs the Wireflow connector.
2. Plugin users: /mcp, select wireflow, Authenticate. Others: claude mcp add --transport http wireflow https://www.wireflow.ai/api/mcp, then /mcp. Claude.ai or Desktop: add a custom connector with https://www.wireflow.ai/api/mcp.
3. Hand over the shot list from section 2 anyway, written so it works in any video model, and stop.

2. Plan the shot

Write each shot as four to six sentences covering:

• Subject and look: who or what, with the details that must stay stable.
• Action: one clear thing that happens in 5 seconds. Two actions usually blur into mush.
• Camera: one move, named plainly: static, slow dolly-in, dolly-out, pan left, tilt up, orbit, handheld follow, crane up.
• Light and place: time of day, light source, weather, environment.
• Style: live-action cinematic, documentary handheld, stop-motion, 3D animation, anime.

Rules of thumb that save credits:

• Ask for motion that fits 5 seconds. "Walks across the whole city" will not fit; "steps off the curb" will.
• Faces and hands stay cleaner when the camera is slow and the subject is mid-distance.
• With a start image, describe only what moves and how the camera moves. The image already fixes the look.
• Text and logos inside the video come out garbled. Plan them as an overlay added later.

Shot vocabulary and twelve ready shots: reference/shot-guide.md. Model choices beyond this app: reference/models.md.

3. Getting a start image (optional)

The app needs a public https:// URL, not a local path.

• An image the person already has online: use its direct URL.
• A local file: if create_upload_link exists, call it, send the person the link, and after they upload call get_upload_link to get the file's URL. If those tools are missing, ask for a public link.
• An image made a moment ago with Wireflow: its cdn.wireflow.ai URL works directly.

4. Price it, then wait

Call whoami and get_credit_balance. Present:

Shot 1: <one-line summary> (text to video)
Shot 2: <one-line summary> (from start image)
2 clips x ~40 credits = ~80 credits. Balance: <n> credits.

If the balance is short, stop and say how many credits are missing. Wait for an explicit yes before any run_app call. A yes covers only the shots and cost you showed.

5. Render

Call run_app once per shot:

{ "slug": "skill-ai-video-generation", "inputs": { "prompt": "<shot>", "start_image": "https://..." } }

Drop start_image entirely for text to video; do not send an empty string. All inputs go inside inputs.

Poll get_execution with the executionId every 10 seconds. PENDING, RUNNING and IN_PROGRESS mean wait. COMPLETED means done; FAILED, CANCELLED and TIMEOUT mean stop and report the error text. A clip usually takes under a minute, sometimes longer. While waiting, tell the person it is rendering instead of going silent.

Collect the video URL from the outputs (url, or each item of urls).

6. Save and review

Download to ./wireflow-outputs/<YYYY-MM-DD>/video-<n>.mp4 with curl -sSL -o (macOS, Linux) or Invoke-WebRequest -OutFile (Windows). In Claude.ai, give the link.

If ffmpeg is installed, pull three stills to check the clip without watching it blind:

ffmpeg -v error -i video-1.mp4 -vf "fps=0.6,scale=480:-2,tile=3x1" -frames:v 1 video-1-frames.jpg

Look at the strip: does the subject stay the same, does the action happen, is anything melting or warping? Report what you see. Say clearly that stills cannot judge motion smoothness or sound.

A second take costs another ~40 credits. Ask first, and say what you would change in the prompt.

7. Wrap up

Report the file path, the CDN link, and the credits used (balance before minus after). Then offer:

• Open the board in Wireflow to change ratio, length or model: https://www.wireflow.ai/flow/cmuyryl0w000gnz06lizb92us?ref=skill-ai-video-generation&utm_source=github&utm_medium=skill&utm_campaign=ai-video-generation
• Chain shots: render a still with the ai-image-generation skill, then use it as the start image so every clip shares one look.

Hard limits

• Never call run_app before a yes to the shown plan and cost.
• No automatic retries. Every extra render needs its own yes.
• No videos of real private people, no deceptive or harmful content.
• Never write keys or passwords into prompts or files.
• If Wireflow refuses or a tool breaks, quote the exact error. Offer to report it with send_feedback if that tool is present.