gamussa/reels-producer-skill

Write talking-head scripts and produce Instagram reels and YouTube shorts

1.77x

Quality

97%

Does it follow best practices?

Impact

64%

1.77x

Average score across 3 eval scenarios

Securityby

Low

Low-risk findings worth noting

Dependencies and Degradation

Name: gamussa/reels-producer-skill
Rating: 90.4 (1 reviews)
Author: gamussa

Run the preflight first, every session. It exits non-zero on a missing required tool, and pipeline.py status blocks on the same check.

python3 scripts/check_deps.py --mode music-video   # or --mode talking-head
python3 scripts/check_deps.py --json               # for gating

Required

Tool	Install	Without it
ffmpeg / ffprobe	`brew install ffmpeg`	nothing runs; every stage shells out to it
`.venv`	`python3 -m venv .venv`	macOS refuses system pip (PEP 668)
librosa + soundfile	`.venv/bin/pip install librosa soundfile`	no beat detection (music-video only)

An ffmpeg on PATH is not enough: builds differ in which filters and encoders they carry (Homebrew bottles and CI runners drop libass, libzimg, libfreetype). The preflight asks the build and reports each needed filter and encoder as available, missing, or unknown when the listing could not be read — unknown is never reported as either of the other two. The lists are FILTERS_REQUIRED, FILTERS_OPTIONAL and ENCODERS_REQUIRED at the top of scripts/check_deps.py. No recipe relies on drawtext; labels go through ImageMagick and captions through subtitles (libass).

Optional, and what each actually costs

Never substitute silently. Name the loss and let the user choose.

Missing	What is lost
`auto-editor`	`tighten_vo.py` uses ffmpeg `silencedetect`. Measured equivalent (9.80s vs 9.87s from a 13s take) — still report which ran
`deepFilter`	`clean_audio.py` uses `afftdn`. Audibly weaker on a noisy room; this is a real quality loss
Node 18+	`make_cards.py` cannot run. Build cards as stills and let `normalize_clips.py` apply Ken Burns
`moviepy`	no engine A/B; `render_reel.py` still renders normally
`yt-dlp`	`fetch_music.py` cannot pull audio from a link
`mlx-whisper`	no local transcription. Talking-head reels are watched muted, so captions are not optional there
Gemini API key	no hook-copy drafts, no stylized thumbnails
official D-Log LUT	the bundled LUT is DJI's whitepaper math, an approximation — not the exact per-model transform

The two fakes to refuse by default

--bpm is not beat detection. It lays a fixed metronome grid and never analyses the audio, so cuts land on clock ticks rather than on the track. detect_beats.py warns, stamps "source": "fixed-grid" into beats.json, and build_cut_plan.py repeats the warning. Never present such a plan as beat-synced; if librosa is missing, install it rather than reaching for --bpm.

A camera-conversion LUT is not a grade. The bundled D-Log LUT applies only to dlog-profile clips. On Rec.709 footage it crushes blacks and pushes colour out of gamut.