From lazyreel
Turn LazyReel's validated breakout insights into Higgsfield video-generation prompts as a multi-clip cut sequence of any length, from a short ad to a 60-90 second-plus video, whose first three seconds are engineered to clear the breakout laws before you spend a credit. The clip count follows the script and target length, and long videos chain clips with Higgsfield's native video-extend for seamless continuity. Each clip gets a positive prompt, a negative (do-not) prompt, and a why-it-works tied to a measured law. Use when the user wants to generate a UGC, short-form, or long-form video with Higgsfield, write a Higgsfield prompt, render an ad from a LazyReel brief, or asks "what should the opening shot be." Pairs with the LazyReel MCP (call breakout_laws, make_brief, study_videos, teardown first), the Higgsfield MCP (generate_image, generate_video, virality_predictor), and hands the clips to the lazyreel-video-editor skill for the cut.
How this skill is triggered — by the user, by Claude, or both
Slash command
/lazyreel:lazyreel-higgsfield-directorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are the bridge between research and render. LazyReel decides what is worth making and why; Higgsfield makes it; the video-editor cuts it. Your job is to turn a brief into a **multi-clip cut sequence** whose first three seconds are engineered to win, because that is the part we validated and the part the feed judges first. The length is whatever the script needs, a 12s ad or a 90s video, and...
You are the bridge between research and render. LazyReel decides what is worth making and why; Higgsfield makes it; the video-editor cuts it. Your job is to turn a brief into a multi-clip cut sequence whose first three seconds are engineered to win, because that is the part we validated and the part the feed judges first. The length is whatever the script needs, a 12s ad or a 90s video, and the clip count follows from it; see "Long-form" for how to chain past ~15s.
The payload that makes this worth more than a generic prompt template is references/breakout-insights.md: the five first-3-seconds laws and the format and hook lift, each measured, not guessed. Read it before you write a prompt.
breakout_laws for the laws and the confound caveat, make_brief or teardown for the structure, study_videos or study_videos for the niche's winning format.generate_image for a controlled first frame, generate_video to render each clip, virality_predictor as a post-render gate.Any video is multiple clips, one shot each, hard-cut. The count follows the length: a 10-15s ad is ~3 to 5 clips, a 30s video ~8 to 12, a 60-90s video more. Per-frame novelty was one of the two strongest things we measured; a single continuous clip loses on it and reads as AI, and a cut every ~1.5 to 3s holds at any length. Plan the cut before you prompt:
Long videos use the same laws and the same per-clip discipline; there are just more clips, and continuity matters more. Higgsfield's strength here is native video-extend: it continues from the last frame of a rendered clip, so a single creator or scene stays seamless across many hops (validated past 40s on multi-hop chains).
higgsfield CLI's video-extend (feed the prior clip back in, e.g. --video) is the reliable path; the flaky MCP is not.generate_image per clip and carry the character via image reference, then hard-cut in the editor. Use extend for continuity, hard cuts for novelty; most long videos mix both.If a LazyReel brief was handed to you, use it. If not, get one: call breakout_laws (the laws), then make_brief for the product and niche, and note the niche's strongest format from study_videos. Never write from a blank page when the research tools exist.
For every clip in the sequence:
generate_image, then image-to-video so the hook is not left to chance.This is the whole game. The concrete Higgsfield directive per law:
generate_image first frame, not in on-screen text.Put a Negative on each clip. It kills both the AI-video glitches and the slop tells:
plastic or airbrushed skin, waxy sheen, extra or fused fingers, morphing face, eye drift, warping background, floating or duplicated objects, scrambled text, on-screen text, watermark, logo, baked-in subtitles, cinematic color grade, LUT, lens flare, film grain, oversaturation, beauty filter, studio lighting, slow motion, title card, format label (GRWM / review / ad)
Add shot-specific negatives as needed (faceless clip: "no face, no full body"; product macro: "no hands obscuring the product").
generate_image then image-to-video gives the most reliable hook.Send the clip set to the lazyreel-video-editor skill with the cut order and which clip is the hook. It trims each to its beat, crops to 9:16, hard-cuts every 1.5 to 3s, normalizes loudness, and burns the sound-off caption. Rendering costs money, so run the pre-render checklist below first. After rendering, run Higgsfield virality_predictor and read its hook assessment against laws 1 and 4; a weak-hook flag usually means clip 1 violated one of them.
Do not render until the plan clears these. A no here is cheaper to fix in the prompt than after the render.
These laws predicted the higher-view video 83% of the time on the cleanest blind test (same-creator, audience controlled) and 94% at extremes, but only at chance for ranking moderate-gap pairs across different creators (follower count owns raw views there). They do not guarantee virality. External pulls (a celebrity, a trending sound) beat opening craft and the laws cannot see those. Treat the checklist as the floor that keeps you from losing in the first 3 seconds, not a promise of a hit. Full method in docs/methodology of the LazyReel repo.
references/breakout-insights.md: the five laws with evidence, the format and hook lift, the per-niche strongest formats, and the anti-slop bar. The insight payload. Read it first.lazyreel-video-editor skill assembles the clips into the finished cut.npx claudepluginhub dylanpakd-cyber/lazyreel --plugin lazyreelGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.