# Flux3 Video Directing > Use when writing or critiquing FLUX 3 video prompts. - Skill: `theheavenlyd3mon/flux3-video-directing` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds add theheavenlyd3mon/flux3-video-directing` - Raw SKILL.md: https://api.skillmd.com/api/skills/theheavenlyd3mon/flux3-video-directing/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: theheavenlyd3mon (https://skillmd.com/u/theheavenlyd3mon) - Updated: 2026-09-09 - Page: https://skillmd.com/skills/theheavenlyd3mon/flux3-video-directing --- # FLUX 3 Video Directing ## Trigger Any request to write, improve, storyboard, or critique a FLUX 3 video prompt. This skill owns the CRAFT layer only: prompt anatomy, cinematography language, scene/beat direction, and the example library. ## Companion — mechanics live elsewhere Tool selection, rate limits, polling, keyframe-index math, and contest workflow live in the `video-generation` skill. Load it for generation mechanics; do not restate them here. This skill tells you WHAT to write; that one tells you HOW to run it. (That skill cross-links back here for prompt craft.) ## The prompt anatomy (THE template) The cross-source consensus shape. Write plain prose in this order — the harness is a reasoning model, not a tag encoder, so keyword tricks and word-order hacks do nothing: ``` [Shot type / camera movement] of [subject] in [setting]. [Subject action over time, with timing beats]. [Environmental motion over time]. Audio: [ambience + event-tied SFX + music direction, or "no music"]. Style: [cinematic style, lens, lighting, texture / film stock]. Constraints: [identity lock, no text, no logos, no extra characters, realistic motion]. ``` ### The 9 rules (condensed) 1. **One clip = one shot.** Write a shot direction, not a scene idea: who's in frame, where the camera is, what moves, what changes over time, what's heard. 2. **Timing beats.** FLUX 3 follows timelines: "0–5s: … 5–10s: …", "after three seconds, a red light begins pulsing", "at the 2-second mark…". 3. **Audio section is mandatory** (it's on by default — silence must be asked for). Layer it: ambience + SFX bound to visible events with "as/when" + music direction or explicit "no music". 4. **Camera language + pace.** Use standard terms (push-in, dolly, truck, pan, tilt, orbit/arc, crane, rack focus, handheld, locked-off, whip pan) and add pace + what each phase reveals. For static: "the camera is entirely motionless for the duration; movement occurs only from the subject." 5. **Anti-artifact phrasing is standard practice for video** — see the tension note below. "One true unbroken continuous shot, no cuts, no morphing", "seamless throughout, consistent characters and world", "preserve face". 6. **Iterate one variable at a time.** Keep a reusable style prefix identical across clips for sequence consistency. 7. **Duration: 5–8s default** (establishing/cutaway/insert/reaction/texture). 10–20s only when the subject stays stable and the action has a beginning-middle-end. Don't default to 20s. 8. **Dialogue:** quote exact lines, name a speaker visible on camera, give voice qualities (accent, pace, "Beat." pauses, tone), "one at a time, never overlapping" for multi-speaker — or let the model improvise from a situation. 9. **Split-screen works exceptionally well:** "15-second split-screen, two equal vertical halves, one continuous take, no cuts, perfectly frame-synchronized" + per-half visibility rules + timed beats + matching rules. ## Directing workflow 1. **Brief.** One sentence: subject, mood, the single thing the viewer should feel. Grounding: research anything with a real checkable appearance and put camera-visible specifics (silhouette, materials, colors, era) in the prompt. 2. **Pick beats.** For suspense/reveal/trailer work, choose beats from `references/suspense-and-trailer-structure.md` (8-beat ominous-reveal table + trailer templates). One beat = one clip. 3. **Write per-clip prompts** with the anatomy above. Pull shot/movement/lighting terms from `references/cinematography-vocabulary.md`; steal proven phrasing from `references/prompt-library.md`. 4. **Lock a reusable style prefix** (e.g. "hyper realistic blockbuster cinematic, anamorphic 35mm, heavy film grain, deep blacks") and repeat it verbatim in every clip + continuation so the world never resets. 5. **Generate** via the `video-generation` skill's mechanics. Two principles to hold throughout: - **Withhold ratio < 15%.** The subject is visible for under 15% of total runtime; everything else is evidence and reaction. Imagination out-scares pixels. - **Sound precedes image.** The threat exists in audio before it exists visually (wing beats, distant roar, stone cracking). Let the next shot's sound enter before its picture (J-cut) to stitch clips seamlessly. ## The negative-prompt tension (read this) BFL's FLUX.2 **image** docs say "no negative prompts — describe what you want." FLUX 3 **video** power users consistently use explicit prohibitions successfully ("no morphing, no timing offset, no extra characters"). For video, the constraint line wins: treat "no X" prohibitions as standard practice. Both can coexist — phrase the positive goal AND forbid the specific artifact. ## Failure modes (top 8) + fixes | Failure | Fix | |---|---| | Invented/unwanted music | Audio is on by default — always write an Audio: section; say "no music" explicitly | | Quoted dialogue burns in as text/subtitles | Make a speaker visible on camera; describe them; add "no on-screen text, no subtitles" | | Multi-shot prompt blends into one take | Consecutive shots must contrast in scale, location, or color | | Morphing/drift mid-clip | "one true unbroken continuous shot, no cuts, no morphing"; strong refs; identical style prefix; explicit physics + timing | | Character inconsistency across clips | Turnaround sheets; reference everywhere; video refs > stills; identical style prefix | | Split-screen / multi-view desync | "perfectly frame-synchronized", per-half visibility rules, matching rules, "no timing offset" | | Overlong clip loses coherence | 5–8s unless the action has a full arc and a stable subject | | Vague one-line prompt → generic output | Director-style specificity: fill all six anatomy slots (camera + subject motion + environment motion + timing + audio + style + constraints) | ## References - `references/cinematography-vocabulary.md` — lookup tables: 15 shot types, 16 movements, 14 lighting terms, 10 lens/DOF terms, each with emotional effect. - `references/prompt-library.md` — verbatim proven prompts (flick.art, X power users, fluxproai templates, Runway/Sora/Kling/Veo fragments), each annotated with why it works. Keep source attributions. - `references/suspense-and-trailer-structure.md` — 8-beat ominous-reveal table, Derek Lieu 4-act + 30–60s teaser templates sized for chained FLUX clips, pacing mechanics. Makes a creature teaser repeatable for any creature. ## Pitfalls - Don't write a scene idea — write a shot. "A scary forest" is not a prompt. - Don't leave text/logos ambiguous: explicitly request (quoted, placed) or forbid ("no readable text, no logos"). - Don't pile unrelated sounds — audio mush. One ambience + event-tied SFX + optional score. - Don't confirm a full reveal for suspense work — user taste is restraint with ONE clear glimpse held for a beat, then gone. Confirm glimpse level before locking beats. - Don't restate mechanics here; if a rule about rate limits/polling/keyframe indices is needed, that belongs in `video-generation`.