# Shortfilm Desk > Find the shortfilm template your idea already is, learn what that pattern historically > cost in rerolls, and write the five-stage prompt it calls for. Five lanes over one idea, > plus a free browser-side matcher and a seven-rule checker. URL: https://shortfilm-desk.skillsafe.ai/ API: https://shortfilm-desk.skillsafe.ai/api.html Derived from jnMetaCode/ai-shortfilm-prompts (https://github.com/jnMetaCode/ai-shortfilm-prompts), MIT: its five-stage structure, its 26-row template routing table, its seven hard rules, its per-model advice and prompts/index.json. Not affiliated with or endorsed by jnMetaCode, Mx-Shell, or any property named in a substitution table. The reroll figures are the upstream author’s own notes, not measurements taken here. ## What the free matcher does It runs entirely in the browser, on the idea you type, with no account and no credits. Every template carries tags. The matcher scores tag overlap at 10 points a tag, then adds 8 if the template’s own length suits the ask and 6 if single-shot or multi-shot agrees. A template with no tag hit does not appear at all - length and shape are tie-breakers between plausible answers, not reasons to be one. Anything scoring under a quarter of the leader is dropped, because a shortlist that includes noise is not a shortlist. The breakdown travels with every row: a ranking nobody can audit is an opinion with a number next to it. ## The template library - `15s-transformation` - 15s single-shot transformation. Single shot, 15s. Needs: the trigger gesture, and what stays imperfect after the change. - `multi-shot-narrative` - Multi-shot edited narrative. Up to 8 shots. Needs: the subject registry and the atmosphere lock, before shot one. - `pet-lifetime-narrative` - Emotional narrative - family, pet, farewell. Up to 8 shots. Needs: two imperfection anchors per subject, and one locked grade across seasons. - `product-commercial` - Product commercial or hero ad. Up to 6 shots. Needs: the single claim, and what the camera can actually show of it. - `food-asmr` - Food ASMR or sensory close-up. Single shot, 10s. Needs: the sound events, in order, with the moment each one lands. - `animal-vlog` - Talking-animal vlog. Single shot, 10s. Needs: the lines, and whether the mouth should move with them. - `movie-trailer` - Cinematic teaser trailer. Up to 8 shots. Needs: the escalation - what each shot raises that the last one did not. - `cyberpunk-city` - Cyberpunk city or atmospheric environment. Single shot, 15s. Needs: what moves in the frame - a static city is a photograph. - `claymation` - Stop-motion or claymation. Single shot, 10s. Needs: the frame rate feel, and the fact that it breaks the breathing rule. - `nature-timelapse` - Nature or landscape timelapse. Single shot, 15s. Needs: the compression ratio, and the one grade that holds across it. - `found-footage-horror` - CCTV or found-footage horror. Single shot, 15s. Needs: the camera's own defects, and that it breaks the breathing rule. - `anime-to-real` - Anime or 2D to live-action. Single shot, 10s. Needs: the trait bundle to keep, minus anything that names the source. - `music-video` - Music video or performance. Up to 8 shots. Needs: the beat map - music is wanted here, which is the exception. - `sports-slowmo` - High-speed slow-motion sports. Single shot, 10s. Needs: the decisive moment, and how far either side of it to hold. - `fashion-film` - Fashion film or editorial. Up to 6 shots. Needs: the movement that is the subject - there is no narrative here. - `travel-vlog` - Travel vlog or sense of place. Up to 8 shots. Needs: the three or four places, and what each one is for. - `drone-fpv` - Drone or FPV aerial. Single shot, 15s. Needs: the route - the move is the content here. - `micro-drama` - Vertical micro-drama. Up to 8 shots. Needs: the hook in the first three seconds and the cliffhanger in the last. - `sci-fi-space` - Hard sci-fi space or zero-G. Single shot, 15s. Needs: the physics - weightless motion, and silence where there is no air. - `car-commercial` - Car commercial. Up to 6 shots. Needs: what the reflections are reflecting - the surface is the shot. - `dance` - Dance film. Single shot, 15s. Needs: continuous full-body motion tied to the beat. - `project-planner` - 3+ shot project - lock consistency first. Up to 8 shots. Needs: the subject registry and atmosphere lock, filled in before shot one. - `genre-camera-sop` - How the camera should move, by genre (a reference, not a template). - `camera-move-library` - Camera-move phrasing, 50 moves (a reference, not a template). - `atmosphere-prefabs` - Atmosphere and quality paragraph, by genre (a reference, not a template). - `negative-prompts` - Negative-prompt block and per-model routing (a reference, not a template). ## The worked examples, and what they cost This is the part almost nobody publishes. - **Zombie Scavenger** (multi-shot-narrative, 1 variant, contains-IP-names) - reroll cost `very-high`: ~400 images and 200+ video shots to yield ~40 final clips, project-wide. - **Kamen Rider / Tokusatsu Transformation Series** (transformation, 6 variants, contains-IP-names) - reroll cost `variable`: 2-3 attempts at the low end, 20+ at the high end. - **Kai'Sa Transformation Series** (transformation, 3 variants, contains-IP-names) - reroll cost `unrecorded`: not recorded upstream. - **Weapon Charge and Combat** (charge-combat, 1 variant, contains-IP-names) - reroll cost `very-high`: extremely low one-shot success; rerolled many times and assembled in post. - **Mech Drop** (mech-drop, 1 variant, contains-IP-names) - reroll cost `high`: high breakdown rate; camera and instruction non-compliance common. - **Cyber Wuxia Style Skeleton** (atmospheric, 1 variant, safe) - reroll cost `unrecorded`: a fill-in skeleton rather than a generated piece. ## The seven hard rules 1. Concrete nouns, never praise words. The banned ones and their replacements: - "cinematic" -> name the camera and lens - "simulated IMAX film camera + Panavision C-series 35mm f/4" - "epic" -> name the scale, or the camera that would shoot it - "movie-quality" -> name the film stock and the grade - "stunning" -> delete it, or name the physical effect - "screen edges stretch slightly" - "spectacular" -> delete it and describe what is actually happening - "perfect" -> delete it - and note that Rule 5 wants imperfection anyway - "handsome" -> "a slight furrow of the brow" / "a hint of contempt in the gaze" - "premium-feel" -> "glazed surface gloss" / "metal brushed finish" / "film grain" - "texture-rich" -> name the texture - brushed, glazed, matte, grained - "detail-loaded" -> name three of the details - "4k" -> write concrete visuals - "low-saturation grey-blue base, film grain" - "high-quality" -> delete it; it describes nothing the model can render - "masterpiece" -> delete it - tag spam is an image-model habit 2. Name a real camera and lens: - epic, big-scene: IMAX film camera + Panavision C-series (35mm, f/4) - gritty cyber, hard sci-fi: Sony Venice + Canon K-35 series - Hong Kong noir, wuxia: Kodak 35mm bleach-bypass - commercial portrait: Canon EF 85mm f/1.2 3. The breathing line, verbatim: "Handheld shot. Throughout, maintain an extremely subtle, breath-like camera float to enhance presence." Both qualifiers are load-bearing. 4. The sound line: "Sound: No score. Production audio only." 5. At least 2 imperfections. Too perfect reads as CG. 6. No effects pile-up at the close - no blinding light, explosion, victory pose, leap into the sky or camera blow-out. 7. No named intellectual property, substituted rather than deleted: - Kamen Rider -> a masked tokusatsu rider in a segmented insect-motif suit - Gundam -> a humanoid mobile-suit mech with a v-fin crest - Iron Man -> an atom-punk retro-futurist red-and-gold combat suit - Kai'Sa -> a void-touched marksman in living purple carapace armour - Kaisa -> a void-touched marksman in living purple carapace armour - League of Legends -> a stylised high-fantasy arena aesthetic - Metal Gear -> a stealth-espionage powered exoskeleton - Pacific Rim -> a piloted skyscraper-scale defence mech - Ready Player One -> a stacked-VR retro-futurist world - The Matrix -> a green-cast simulated-reality aesthetic ## The models Each decides three things: what it filters, how long it runs, and where the negative prompt goes - and the third differs three ways, not two. - `seedance` - Seedance 2.0 (Doubao / Jimeng). strict IP filter, zh or en, 4-15s. Negative: a dedicated field. Jimeng web and VolcEngine do 4-15s single-shot, but the Doubao app is locked to 5s and 10s - do not promise 15s on Doubao - `veo` - Veo 3 / 3.1. strict IP filter, en, 8-8s. Negative: a dedicated field. 8s a clip, extended in 7s hops; the negative field takes plain noun phrases, not "no..." commands - `kling` - Kling 2.x / 3.0. strict IP filter, zh or en, 5-15s. Negative: a dedicated field. a pre-generation banned-word filter rejects the WHOLE prompt on one flagged term, so sanitise body and contact words first; 5-10s, 3.0 up to about 15s - `hailuo` - Hailuo / MiniMax. moderate IP filter, zh or en, 5-10s. Negative: a dedicated field. resolution trades against duration - about 6s at 1080p, about 10s at 768p - `wan` - Wan 2.x (Alibaba, open-source). lenient IP filter, zh preferred, 3-15s. Negative: a dedicated field. lenient when self-hosted; leans Chinese, so add ZH for tricky or first-last-frame shots; 3-8s, newer builds 10-15s - `runway` - Runway Gen-4 / 4.5. strict IP filter, en, 5-10s. Negative: no field, and "no X" summons X - describe only what should appear. no negative field, AND "no X" can summon X - describe only what SHOULD appear - `pika` - Pika 2.2 / 2.5. moderate IP filter, en, 5-10s. Negative: a dedicated field. 5s and 10s standard; Pikaframes keyframes reach about 25s but are not general; 2.5 supports negatives, verify 2.2 in-app - `sora` - Sora 2 / 2 Pro. strict-descriptions IP filter, en, 5-25s. Negative: no field, fold into the positive prompt. a triple-layer filter catches lookalike DESCRIPTIONS, not just names - avoid recognisable trait bundles; up to about 25s single-pass on Pro; no negative field, so fold the guardrails into the positive prompt ## The five stages 1. **Core theme** - 3-6 tags separated by | , ramping from shot type to genre to aesthetic 2. **Character & scene** - three lines - Face, Clothing, Scene - with the reference sentence first 3. **Atmosphere & quality** - a real camera and lens name, and a colour phrase 4. **Camera rules** - three lines - single-shot or edited, angle, and the breathing sentence 5. **Storyboard** - per-second slices (Action + Camera + VFX) or per-shot slices (Size + Composition + Move + Action) ## The five lanes ### `match` - Find the template Twenty-six genre templates and six worked examples, ranked against your idea with the arithmetic shown - and, where the library recorded it, how many rerolls that pattern actually took before it produced a usable clip. Sections: `Summary`, `Shortlist`, `What It Needs`, `Cost Note`, `Next Step`. Fields: `idea`*, `seconds`, `shots`, `model`, `language`. ### `plan` - Five-stage prompt The prompt itself, in the five-stage structure: core theme, character and scene, atmosphere and quality, camera rules, storyboard. The seven hard rules are checked in your browser before you spend anything. Sections: `Summary`, `Prompt`, `Rule Check`, `Model Note`, `Next Step`. Fields: `idea`*, `template`, `seconds`*, `model`*, `subject`, `draft`, `language`. ### `board` - Storyboard slices Stage five on its own, at the depth a shoot needs: per-second slices for a one-take, or per-shot slices for an edit - and never more than eight shots. Sections: `Summary`, `Storyboard`, `Slice Table`, `Restraint Check`, `Next Step`. Fields: `prompt`*, `style`, `seconds`*, `shots`, `language`. ### `safe` - IP and model routing Every model filters differently, and one flagged term can reject a whole prompt before generation starts. This rewrites the named properties out and routes the negative prompt to where that model actually reads it. Sections: `Summary`, `Named Properties`, `Rewritten Prompt`, `Negative Routing`, `Next Step`. Fields: `prompt`*, `model`*, `seconds`, `insist`. ### `audit` - Seven-rule audit Paste a prompt you already have. Every finding names the rule and quotes the exact line to add, and the lane returns a repaired prompt rather than a list of complaints. Sections: `Summary`, `Verdict`, `Findings`, `Repaired Prompt`, `Next Step`. Fields: `prompt`*, `model`*, `seconds`, `focus`. ## What this app does not do It writes prompts and plans. It does not generate video, it calls no generation API on your behalf, and the reroll figures it reports are the upstream library’s notes rather than anything measured here. The prompts are yours to paste into Seedance, Kling, Veo, Sora, Runway, Pika, Hailuo or Wan.