Feverdream

Use this skill to create Feverdream edits: strange found social-video memories built from real footage, object-mask transitions, compressed color, and loopable dream logic.

This works like black-mirror as a practical editing system, but the style is less cursed monitor-wall and more old-internet dream loop.

Core Rule

Do not immediately imitate or render. First inspect the reference, describe the style clearly, and ask the user to confirm. Only after confirmation should the style be used for an edit or saved as a reusable style skill.

If the user directly asks for feverdream with no new reference, use the confirmed default style:

Old internet dream loop with object-mask transitions:
surreal real-world clips, bright compressed footage, foreground objects
used as wipes, cuts hidden at peak object coverage, minimal effects,
small social-video framing, loopable ending.

Workflow

1. Accept the reference video path or uploaded file.
2. Inspect technical metadata: duration, frame rate, resolution, audio presence, and scene/cut density.
3. Generate frame samples/contact sheets at multiple rates.
4. Measure rough cut rhythm with scene detection when possible.
5. Read the edit style from frames, timing, motion, audio relationship, and platform context.
6. Summarize the proposed style in plain language.
7. Ask the user to confirm or correct the style.
8. After confirmation, build the edit recipe and apply it to the requested source video.
9. QC the output against the reference style and adjust if it drifts.

Fast Analysis Command

Use the bundled helper when a local reference video is available:

python3 scripts/analyze_reference_video.py ./reference.mp4 --out-dir ./reference_analysis

The helper creates metadata, contact sheets, and detected scene-change timestamps. Use these as evidence for the style read; do not treat them as the whole answer.

What To Analyze

Read references/style-analysis.md when analyzing a new reference.

Always look for:

• Cut rhythm: hard cuts, jump cuts, beat cuts, long holds, stutters, flashes, match cuts.
• Mask transitions: foreground-object wipes, occlusion cuts, shape reveals, hand/object covers, and match cuts hidden by motion.
• Layout: full-frame, split-screen, grids, picture-in-picture, overlays, borders, crop behavior.
• Motion: handheld, zooms, pans, shake, freeze frames, speed ramps, reversed moments.
• Color: contrast, saturation, tint, black levels, bloom, grain, halation, VHS/CRT texture.
• Text/graphics: captions, timestamps, labels, UI marks, subtitles, logos, stickers, symbols.
• Source feel: phone footage, archive footage, music video, surveillance, TV broadcast, screen recording.
• Audio relationship: cut on beat, off-beat disruption, silence, risers, impacts, tape stops, voice sync.
• Repetition: recurring transitions, motifs, framing tricks, panel logic, glitch language.

Senior Read

Think like an editor, not an effect preset.

• Separate container from content: platform chrome, aspect ratio, caption layout, and compression are part of the style if they define the viewing experience.
• Separate cut rhythm from camera motion: a clip can feel fast because of internal motion even when cuts are slow.
• Separate grade from source quality: old compression, phone exposure, archive transfer, and intentional color treatment are different signals.
• Identify the editorial promise: dream memory, surveillance, music-video performance, documentary proof, hype montage, comedy timing, luxury polish, horror unease, or social repost artifact.
• Name the "do not cross" line: what would make the edit stop feeling like the reference.
• When music is present, decide whether visuals obey the beat, fight the beat, or float over it.
• Treat transition mechanics as style-critical. If a reference hides cuts behind objects, hands, signs, doors, cars, camera passes, or frame-filling foreground shapes, call that out as a mask-cut system instead of labeling it a simple hard cut.

Confirmation Format

Before editing, respond with a compact style brief:

I read the reference as:
- Pacing:
- Frame language:
- Color/texture:
- Transitions:
- Audio behavior:
- Things to avoid:

Confirm this and I’ll use it as the edit style.

If the user confirms, treat that summary as the temporary style contract for the job.

Rendering Guidance

Prefer real editing operations over vague generation:

• Use ffmpeg for cuts, crops, stacks, overlays, speed, grade, grain, text, and final render.
• For mask cuts, build transitions around frame-filling motion: cut when an object covers the lens, use wipes that follow foreground shapes, or match one large object/color block into the next shot.
• Use the local black-mirror renderer as a toolbox when the confirmed reference calls for split screens, tape texture, archive-feel, grids, PIP, CRT, channel-zap, or collage behavior.
• Use contact sheets for QC before delivering.
• Preserve the user's source footage unless they explicitly ask for generated footage.
• If reference style depends on copyrighted footage, learn only the editing technique and visual language; do not copy protected characters, logos, or exact shots.
• If the reference includes a real person's identity or private content, do not identify them. Analyze only the edit, camera, timing, layout, and visual treatment.

Render Recipe (v1)

The bundled scripts/feverdream_edit.py is the canonical Feverdream renderer once a style is confirmed. It encodes the mask-cut + dream-loop language as a single ffmpeg pipeline with optional song + beat-aligned cuts.

# Default — natural source audio, speed ramps on, 1920x1080, ~18s
python3 scripts/feverdream_edit.py --output ./fd.mp4

# Drop a song to replace audio
python3 scripts/feverdream_edit.py --output ./fd.mp4 --audio ./song.mp3

# Beat-cut: numpy onset detection re-times segments to land cuts on beats
python3 scripts/feverdream_edit.py --output ./fd.mp4 --audio ./song.mp3 --beat

# Disable speed ramps if the confirmed style calls for raw cuts
python3 scripts/feverdream_edit.py --output ./fd.mp4 --no-ramps

What the recipe applies:

• Speed ramps with motion blur per segment — slow front (0.7×) → fast back (1.7×), joined by tmix=frames=4 shutter persistence. Source rendered internally at 60fps so the blend reads as organic motion blur.
• 12-transition palette — fadewhite, pixelize, hlslice, fadeblack, dissolve, smoothleft, radial, distance, vuslice, smoothup, hrslice, vdslice cycled per cut.
• Pre-cut accents — every outgoing segment's last 0.18s gets a stab: punch_zoom (scale 1.0→1.18), hue_stab (sat 1.9 + +20° hue), negate_flash (last 2 frames inverted), chroma_shift (psychedelic hue ramp). Cycled per transition.
• Per-segment grade — bright compressed (sat 1.45–1.60, contrast 1.10–1.22, gamma 1.05–1.10), with hue shifts and a soft dream-blur on one segment.
• Bloom blend — duplicate-blur-screen at 22% for the "old internet bright compressed" feel.
• Beat detector (no external libs) — pure numpy onset detection. Extracts mono PCM via ffmpeg, frames to 1024-sample windows, computes positive energy flux, peak-picks against an adaptive local mean (1.6× threshold), enforces 0.28s minimum gap.

Edit SEGMENTS at the top of the script to change the source roster or per-segment offsets. Default sources point to a Prelinger / public-domain pool cached at ~/black_mirror_work/sources/ (the black-mirror skill's cache).

Output Standard

Deliver:

• Final video path.
• Short note explaining which confirmed style traits were applied.
• Any limitations, such as missing audio beats, low source quality, or effects that need a second pass.

Keep the experience practical and fast: inspect, confirm, edit, QC, deliver.