Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Presentations

film-taste

The taste baseline for every edit and motion film in this kit, so the work looks made by a person and not by a template or an AI tool. Load it before planning, building or reviewing any video, and whenever someone says a cut looks "AI", "stock", "template", "like slides", "cheap" or "PowerPoint".

edotheDev · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add edotheDev/claude-video-kit --skill film-taste

About this skill


name: film-taste description: The taste baseline for every edit and motion film in this kit, so the work looks made by a person and not by a template or an AI tool. Load it before planning, building or reviewing any video, and whenever someone says a cut looks "AI", "stock", "template", "like slides", "cheap" or "PowerPoint".

Film taste

The baseline is neutral and human: real footage, real type, real sound, cut on purpose. Nothing on screen should look like a default someone forgot to change.

The one test. For every choice, ask whether a working editor would have made it on purpose. If the true answer is "the tool did it", change it.

Footage

  • Real before generated, specific before generic. A shot of the actual thing (the real company, place, person, document or product) beats any illustration of it.
  • The stock look is the first tell. Posed models acting out a feeling, hands typing on a white desk, glass meeting rooms, rim lit portraits, a phone on a tripod in an empty office. They read as filler even from a free library and even with a film grade on top.
  • What reads as real: candid moments, film scans (8 mm, 16 mm, Super 8), public domain archive film, nature, science and microscopy, official uploads from the subject itself, real interviews and news, screen recordings of the real product.
  • Mix sources. One library makes every film look alike.
  • Full frame, natural colour. Unify with a light grade and fine grain, never a heavy LUT. A brand palette is for type, grounds and the end card, not for repainting footage.
  • Each source moment appears once. Count by camera setup, not by file: the same stage from a new in point or a new crop is the same shot. A longer voice means sourcing more clips, not re-cutting the ones you have.
  • HD is the floor. Do not blow up SD archive to fill a vertical frame; window it so it stays sharp.
  • Clean the frame. Crop out watermarks, burned in captions, creator handles, laurel cards and third party logos. Check the final frames, not just the sources.

Type on screen

  • One face per position for the whole film. For example, centred lines in one face and bottom lines in another. Never swap faces from screen to screen and never mix faces inside a line.
  • One size per screen, no italic beside roman. Emphasis comes from the picture, a pause or a held frame, not from a second size or style.
  • Words appear as they are said: a short fade up out of a slight blur, timed to the voice. No bounce, no pop, no scramble, no typewriter at a constant rate, no karaoke colour sweep.
  • Text never jumps. If a face sits under the words, reframe the footage, do not move the words.
  • Legibility without boxes. Dim the shot under the words (20 to 45 percent) and use one soft shadow. Captions on a talking head get a thin outline, no backing box.
  • Safe areas. Keep words and logos clear of the platform UI: nothing important in the bottom fifth of a vertical frame or along the right edge.
  • Copy reads like speech. Short, plain, spoken lines. No hype words ("seamless", "effortless", "supercharge", "game changer", "AI powered").

Pace and cuts

  • Cut on the beat of the music that ships. Pace follows the music: short bursts where the song hits, about one cut a second between, longer holds where it thins or a headline is read.
  • Never a uniform rate. Two identical actions never take identical time.
  • Hard cuts by default. A transition has to mean something: a match cut, a dropout to black before the turn. Whoosh, slide and zoom transitions on every cut are the template look.
  • Stillness is a move. A held frame after a hit and a beat of black before a twist land harder than any effect.
  • Vary the structure. If a viewer who saw the last two videos can predict this one, change its spine.
  • Land the ending. The voice finishes, the picture holds, the music plays on and fades over 2 to 3 s. Never a dead stop.

Sound

  • Music under a voice is instrumental. Check with a stem split, never from memory or a track's title.
  • One track, continuous, chosen for this film. Say why it fits. A commercial song is a reference for rhythm, never the track.
  • Sound effects are the exception, not the texture. Whooshes, risers and cinematic hits on every move fight the music. In a narrated piece, at most a soft click where a photo lands.
  • Leave a real voice alone. Cut it, fade each join, set one gain. No denoise, EQ or compression chain unless asked.
  • Measure loudness. About -14 LUFS integrated, true peak at or under -1 dBTP.

Colour, texture and graphics

  • Banned: blurred gradient blobs and aurora backgrounds, violet to indigo ramps, gradient text, lens flares and light leaks as punctuation, glow on type, glassmorphism, rainbow status chips, sparkle icons for anything AI, floating 3D UI mockups, a fake cursor gliding in perfect arcs.
  • Graphics show something true: a number counting to its real value, a highlight bar on a real headline, a tracking box on a real object. Never invented UI or invented numbers.
  • Texture unifies. Fine grain and a light vignette, nothing heavier.
  • Impact frames (flashes, negatives, glitch slices, stutters) belong to hype edits, where they are the style. Keep to 3 full frame flashes or fewer in any second.

The slideshow trap

A video looks like slides when it is a still, a centred title and a fade, repeated; a slow zoom on every photo; bullet points on screen; title cards between sections; sixty seconds of symmetric centred layouts; a logo sting at the start.

The fix: video over stills; the voice carries the information and the picture carries the proof; one idea per screen; words appear as they are spoken instead of sitting there as a slide; vary the framing; photos only where a group of pictures tells the beat better than video, laid flat and landing one per beat.

AI generation

  • Generate only what ships. One attempt
Read the complete source ↗