This is an 11 minute recording that details how Claude Code edits my videos, including the cuts, the shorts, the generated backgrounds, lizards crawling on the walls, a robot arm, and one small fire.
Watch it here: https://youtu.be/wCdnHdamSt4
Why I do this
Once a week I record a horizontal video, on my own or with my accountability partners. That one recording becomes YouTube long form, shorts for Instagram, YouTube Shorts and TikTok, teasers for a paid course or community, and clips to embed on a website. Recording is the part only I can do. The editing is the part I handed to Claude.
If you run a business and record videos, or keep meaning to start, this pipeline can help you start the process. You need video content and a Claude subscription, or another powerful AI, plus a few starter Higgsfield credits. That is the whole entry ticket.
The setup
Claude: the $100 a month subscription I pay for anyway. Editing videos is a bonus on top, and it does not use extra credits.
Higgsfield: the only video-specific cost. $60 a month for 1,200 credits, priced at $30 the first month. About 20 finished shorts per plan after you learn the setup.
Free local tools Claude drives for me: ffmpeg, Whisper transcription, background removal and matting.
Five things I learned the expensive way
Matte and composite instead of transforming footage. When a video model redraws a frame, it redraws your face. Cut yourself out locally for free, generate only the background, and your face is camera pixels, no generation needed.
Generate a still first. A 3 credit still image settles the wardrobe, the lighting and the composition before spending 20 to 120 credits on a small video. My costume change landed on the first take because the still was approved first.
Model choice is paramount and can be expensive. Same prompt, same photo: NanoBanana and Flux returned a stranger. Seedream 5.0 Pro returned me, including AirPods, my hair, my hands and all.
Generate background effects, loop locally. Two 10 second plates became a 55 second world behind my talking head, and the loop reuses across a whole series for free.
Resolution and mode are the cost lever. The same shot runs 21 credits at 480p fast and 126 at 1080p standard. Shorts play on phones. Cheap tiers look the same on a phone.
The receipts
Portrait (vertical) background: Finished maze short: 73 credits, $3.65
Landscape (horizontal) background: Finished room swap and costume change: 52 credits, $2.60
Animated elements: 3D Claude mascot, one look-check still plus two animations: 45 credits, $2.25
Various: Takes that reached no finished file: 158.5 credits, $7.93
Everything through the day I recorded: 370.5 credits, $18.53
VFX effects added to the intro, a robot arm and lizards (twice): 87.5 credits, $4.38
All-in spend: 458 credits, $22.90 (close to a third of my $60 Higgsfield investment) - including a reusable shorts animation.
Almost half the early spend went to takes nobody saw. That waste bought the knowledge: Claude does every free step first, prices every generation, and I approve paid steps one by one during setup. That generates predictable costs and output expectations.
The prompts
The golden rule behind all of them: preserve everything, change one thing. Every working prompt opens by locking what must not change, and only then names the single change.
Face-lock still (Seedream 5.0 Pro): Change only the clothing, the room behind her, and the lighting. Do not change her face. FACE, ABSOLUTE PRIORITY: her face, head and hair remain the original photograph: same structure, same eyes, same age, same mid-speech open mouth, same AirPod. Do not redraw, reshape, slim or replace her face. Only the light falling on it changes. Then describe the one change in full detail.
Mascot look-check still (3 credits before any animation): Turn this flat pixel mascot into a physical 3D voxel figure, keeping its exact design: same blocky proportions, same flat orange color, same rectangular eyes. The figure sits at a tiny open laptop, hands resting on the keyboard. Matte 3D-printed plastic, soft warm key light, dark background. Photoreal render. No text anywhere.
Mascot animation (from the approved still): Animate this 3D-printed voxel figure photograph. Preserve exactly: the design, the material, the laptop, the background, the camera. Camera locked. Motion: the blocky hands type continuously with quick believable key presses; the screen casts a soft flickering glow; the final pose matches the opening pose for a seamless loop.
Video-to-video skeleton (the robot arm and the lizards both use it): Source: describe the real clip precisely, wardrobe, room, camera, light. Preserve identity, face, expression, lip-sync, wardrobe, gestures, framing and camera exactly. Change only X. Then stage the change step by step with hard boundaries, state what the effect may never touch, and end with: SFX only, no voice, no source dialogue.
Attach a reference image for anything that must look specific. Every failure I paid for came from a missing reference or from asking for two changes at once.
Do it yourself
The free library at marketcraftinstitute.com collects these prompts and the rest of my working set. And if you want to see what a non-coder builds with Claude beyond video: the Job Signals Report at b2bmarketcraft.com decodes any job posting, detailing what a hiring team really wants. I vibe-coded it start to finish.
Talk soon,