dramake

Dramake

Treat short drama production as a continuous production line, not a pile of prompts. Let users without screenwriting, directing, or editing experience also start from a piece of text and complete story, script, storyboard, generation, voiceover, editing, and quality check under clear acceptance gates; let professional users take over from any stage.

This Skill is not bound to a single platform, website, API, or MCP; the execution platform is chosen by the user. Use models and execution entry points that the user can legally access and is willing to use; before each execution, verify the current parameters, prices, and material upload and storage rules.

Use project-state.md as the sole production state machine. Prompts, interface parameters, and timelines are all derived from the approved creative state and the inspected real materials.

Ten Iron Rules

The protagonist must causally change the outcome; people, animals, products, or objects cannot be mere decoration.
Watching muted should also make the main causality and key punchlines understandable; dialogue, subtitles, and extras' laughter cannot replace visible setup, cause, reaction, and reveal.
When the model supports it, prefer fewer, longer story segments, but the event load must be limited: a segment may contain multiple beats, but one prompt must not simultaneously carry multiple high-risk physical chains, prop variants, and unrelated tasks.
Without actually downloading and accepting upstream materials, downstream segments that depend on their state must not be generated.
Lock identity, material, clothing structure, axis, props, weather, humidity, lighting, and sound boundaries.
Contact must have approach, contact, force application, response, and a stable result; reject clipping and impossible anatomy.
By default, do not generate phone UI, letters, price tags, countdowns, and key text in the video; overlay them in post.
Native dialogue can solve lip sync, but each line must lock a unique speaker, visible mouth shapes, and the closed-mouth state of non-speakers; native music cannot automatically solve final-film mixing.
Technical cleanliness does not equal a publishable story; a 60–120 second story must have a macro causal chain running through the whole film and a foreshadowed main twist, and must undergo a cold-viewer check with an unread synopsis.
Without actually watching, listening, or checking frames, you must not claim that the material has passed.

Start from what the user knows

First judge the depth of the user's input; do not require beginners to learn professional terminology first:

**One sentence/brainstorm:** proactively provide 2–3 shootable directions, filling in protagonist, desire, obstacle, escalation, and result;
**Story/novel/copy:** extract characters, relationships, timeline, props, secrets, twists, and visualizable actions;
**Script:** check causality, setup, twist, dialogue density, number of scenes, and duration;
**Storyboard/prompt:** check beats, continuity, physics, text risk, and model executability;
**Existing material/rough cut:** actually inspect picture and sound, locate the smallest failing unit, and do not tear everything down and start over without reason.

Then choose the task depth:

**Diagnosis/advice:** analyze only, do not generate, do not pay;
**Quick fix:** fix only the smallest unit of script, prompt, shot, seam, voice line, or mix;
**Production package:** Output brief, script, master, storyboard, checklist, prompts, costs, and editing plan;
**Execution/continuation:** Continue from the currently accepted state;
**Full production:** Execute the entire state machine from text to publishing.

Local issues must not be forced to go through the entire process again.

Choose content format and aspect ratio

First choose the primary format for the current delivery, see format-routing.md：

**15–45 second standalone short video:** one hook, one escalation, one payoff;
**45–120 second standalone short film:** 2–4 story segments, completing an emotional or comedic arc;
**1–4 minute episodic series:** fixed characters/world rules, with each episode's problem resolved independently;
**1–4 minute serialized short drama:** in addition to the episode goal and payoff, also maintain season arc state.

Supports 9:16、16:9、1:1 and custom ratios. Record the publishing channel, aspect ratio, subject safe area, subtitle area, cover crop area, and delivery encoding.

When both horizontal and vertical versions are needed, share the story, characters, voices, props, and continuity master, while designing composition, shot size, movement paths, and safe areas separately, and accepting them separately. Do not default to center cropping or mechanical edge filling.

Create a production brief

Use production-brief.md. Infer safe defaults from the conversation and materials, and only ask about choices that would materially change the result.

Record at least:

the original text and adaptation boundaries;
channel, aspect ratio, duration, audience, language, and emotional promise;
characters, animals, products, locations, reference materials, and usage rights;
standalone/episodic/serialized format and content configuration;
available models/execution entry points and approximate budget ceiling;
dialogue strategy: no dialogue, native, reference audio, external voiceover, or hybrid;
visual medium: live action, realistic pets, animation, motion comic, UGC, or hybrid.

Public examples must not expose keys, signed links, request IDs, private paths, real faces/voices, or private reference materials.

Write the story first, then the prompts

Read story-and-pacing.md and directing-camera-and-pacing.md; for comedy, cute pets, anthropomorphic, or animated performance projects, also read comedy-and-animated-performance.md. Deliverables:

a one-sentence audience promise;
a 0–3 second visual hook;
a motivation unique to that protagonist;
a visible goal and obstacle;
a change in information, action, reaction, relationship, or risk every 4–8 seconds;
a relationship or state change;
a payoff/twist with setup;
a final emotional beat, laugh, suspense point, or character-defining action.

For 60–120 second comedy, first write 触发事件 → 主角误解/欲望 → 目标 → 主动选择 → 代价升级 → 主反转 → 后果/回收, then place local jokes. Triggers should preferably be experienced firsthand by the character in the flow of life, rather than standing still at the opening and listening to a full explanation; every joke must change the mainline state and be accepted according to 铺垫 → 可见原因 → 观众信息优势 → 反应停顿 → 揭晓 → 后果 .

Define for each story segment entry_state, observable beats, and exit_state. A 20–30 second long segment usually has 4–6 beats; a 10–12 second dialogue segment usually has 2–3 beats. Continuously add breathing, blinking, gaze, expressions, species-appropriate reactions, prop interactions, and background life; do not let the character become still after finishing a line.

When the goal is 60–120 second Douyin high-density drama, enable the corresponding preset: a clear conflict appears within 3 seconds, the immediate goal is explained within 10 seconds, 90 seconds starts with 18–24 effective beats, and one micro-task is closed every 15–25 seconds; a roughly 30-second continuous voiceover scene includes multiple rounds of advancing dialogue and 4–7 visible actions/reactions. One fixed macro goal runs through the whole film, and scene changes must have a causal bridge. Effective beats are not equal to cuts or generation requests, and must not be re-fragmented and forced together.

If any of the following occurs, go back for revision before generation:

The plot remains unchanged if the protagonist is deleted;
The motivation is explained only through dialogue;
The twist has no visible clues planted in advance;
Characters repeatedly express the same meaning;
Location, time, goal, or character knowledge suddenly changes;
A cold viewer cannot state the goal, obstacle, choice, and result.

Lock content configuration, characters, style, and world

Read content-profiles.md, choose one primary configuration, and stack specialized rules when necessary.

Create a master and character-bible.mdfor each recurring character/animal, recording identity, anatomy, face/body proportions, markings, costume structure, allowed behavior, prohibited variation, and voice. Use identity-voice-and-reference-locks.md to specify a single authoritative reference for each control dimension, and establish an independent voice anchor for each recurring speaker.

Before storyboarding, inventory all props that repeat, cross scenes, are plot-critical, or carry text/Logo, and create a prop-bible.mdfor each. First use the user-selected and verified high-fidelity image generation/editing route (which may include GPT Image-type models) to lock neutral-light multi-view geometry, scale, materials, connectors, and state; create a separate text_plate_idfor precise text, with post-production tracking compositing by default. Every shot references the same prop_id + state_id + text_plate_id; scales, seals, helmets, phones, or packaging must not be redesigned across acts.

Live-action projects lock real photography, skin/fur, actual costume materials, real scale, motivated lighting, and physically integrated effects. Animated projects lock design grammar, lines, shading, proportions, and motion language. Pets in live-action environments may choose natural realism, enhanced realism, or “animated actor hybrid”: the environment and materials are realistic, while the character uses locked anthropomorphic poses, an expression library, anticipation, overshoot, settle, and controlled squash/stretch; props and contact still obey real continuity. “Animated exaggeration” must not be used as an excuse for clipping, teleporting, or prop drift. Read style-and-medium-locks.md and comedy-and-animated-performance.md。

Use continuity-ledger.csv to maintain official state, updating it for every acceptance segment.

For long novels, serial dramas, or cross-episode IP, also read series-memory-and-source-adaptation.md, using layered facts, character knowledge, timelines, and reveal state, rather than repeatedly stuffing the full text into prompts or relying on dialogue memory.

For multi-character staging, real locations, chases, mirrors, giant objects, facilities, complex entrances/exits, or anomalous spaces, also read spatial-blocking-and-world-state.md. First lock location anchors and connectivity relationships, then write character start → path → end, camera positions, sightlines, occlusion, scale, and reflection topology; prompts must not invent positions on the fly.

In high-risk or strongly continuous projects, read previsualization-and-coverage.mdbefore video generation: first make long-term masters for characters/locations/props, then test-shoot a small number of coverage options for key plot moments, and finally freeze only the first frame, last frame, or state frame that will actually constrain the request. Simple reaction shots do not require a complete storyboard; multi-character scenes, giant-object scale, mirror rules, real landmarks, precise blocking, and key contact must not skip previsualization.

When you need to generate character design sheets, scene images, prop images, or storyboard images, also read reference-graph-and-storyboard-images.md. Organize the assets into a dependency graph with responsibilities: first approve one composition master frame, then generate subsequent storyboard shots; for each reference, explicitly specify controls and must_not_transfer, and a multi-pose sheet must not be duplicated into extra characters.

Design generation segments and editing beats

Read generation-blocks-and-edit-beats.md and directing-camera-and-pacing.md. When the model reliably supports it, use 15–30 second segments; use shorter segments for precise dialogue, complex contact, high motion, deformation, or drift-prone tasks.

Each long segment must have:

3–6 timed beats;
one primary action or information change per beat;
2–5 motivated internal cuts/viewpoint changes when needed;
a stable range of time, place, and axis;
the prop_id/state_id/text_plate_id of repeated props and the only allowed variation;
the start point, control input, world/screen vector, phase sequence, and end state of precise physical events;
macro_goal_id, the segment's single advancement of the main line, and the event load budget;
the punchline contract, eyeline/occlusion logic, reaction character, and 0.6–1.0 second reaction pause for comedy segments;
a single impact_role and visible consequence when needed for a hook/collision/reveal/climax;
editing headroom with action still present in the final 0.5–0.8 seconds;
a clear exit action that the next segment can pick up.

For each external seam:

use the approved tail frame when supported, but also provide identity/material masters;
maintain screen direction and action phase;
overlap action by about 0.3–2 seconds when necessary;
transition using action, occlusion, whip pan, foreground pass, sound, eyeline, or shape match;
do not simultaneously change location, action, shot size, direction, costume, and weather without explanation.

See continuity-and-transitions.md。

A long segment is not "cramming more things into one request." Prioritize one main prop system, one high-risk physical chain, and one main comedic/emotional payoff per segment; if multiple locations, food types, precise handoffs, multiple people taking turns speaking, or multiple prop state changes are needed, split the segment or switch to internal establishing/peak/consequence shots.

Formal prompts must be compiled from an approved state; read prompt-compiler.md. Use a stable 场景上下文 → 参考职责 → 空间与首状态 → 镜头 → 时间节拍 → 表演/物理 → 灯光/声音 → MUST-HIT/终态 → 高损失负面约束 skeleton, then trim according to T2V, first-frame, first-and-last-frame, R2V, continuation, editing, or multi-shot mode. Show a full request preview before payment; any prompt, parameter, reference, or schema change that affects the image or cost invalidates the previous confirmation.

Choose models by capability

Read model-router.md, and verify the current entry point before submitting.

**Wan 3-type long segments:** prioritize testing 20–30 second multi-beat segments, reference-driven native dialogue for multiple characters, fewer external seams, and 720p economy/balanced tasks; choose either hybrid R2V or strict first-and-last frames, and focus on checking character/voice binding and drift in the latter half.
**Seedance 2.5-type:** high-value dialogue, multi-image/video/audio references, reference voices, editing, continuation, and repair.
**Seedance 2.0:** 4–15 second dialogue, action, and transitions.
**Seedance 2.0 Fast:** previz, coverage, alternatives, and economy tier; only enter the final film after strict approval.
**MiniMax H3-type:** short high-motion shots, first-and-last-frame control, and key audiovisual shots.
**Kling 3 Omni class:** element binding, multi-shot internal choreography, and first/last keyframe bridging; validate sub-shot payloads and internal cuts first.
**High-fidelity image model:** character/location/prop masters, coverage plans, and frozen keyframes; proceed to motion generation only after the stills pass.

These are only starting assumptions, not permanent rankings. The capabilities actually exposed by the interface outweigh brand impressions.

Choosing quality and paid licensing

Read budget-and-quality.md, and establish a cost plan before paying. Prices are always project configuration, never written as permanent facts.

**Sketch:** minimum viable previz;
**Economy:** usually 720p, long segments prioritized, spot fixes;
**Balanced:** use strong reference/audio models on the segments that genuinely raise the pass rate;
**Premium:** rich multimodality, reference audio, editing/continuation, and selective high resolution;
**Hero Insert:** concentrate budget on hooks, climaxes, complex contact, or emotional close-ups;
**Final Polish:** only upscale and repair already-approved material.

For licensing, separately choose from authorization-and-batch.md in plan-only、single-test、approved-batch or autopilot-with-cap. A quality tier does not imply a paid license. When the user has pre-authorized the platform, upload rules, budget, and retry limits, do not repeatedly ask item by item; when a boundary is reached, you must stop.

For formal execution, also read production-operations-and-recovery.md, and use generation-jobs.csv to store request fingerprints, idempotency keys, task states, dependencies, costs, recovery, and candidate lineage. On timeout, first enter unknown and recover the task; do not treat an unknown state as a failure and blindly resubmit; after manually changing upstream states, invalidate the affected confirmations and downstream tasks.

Choosing dialogue and voice routes

Read audio-and-dubbing.md, and choose per scene rather than blindly unifying across the whole film.

**No dialogue:** unified ambience bed, music structure, and selected foley;
**Native dialogue:** short lines, clear speaker, action while speaking, followed by reaction and the next beat;
**Reference audio:** one authorized voice anchor per recurring character, locking language, pace, timbre, and intensity;
**External dubbing:** use lip-sync tools, or protect credibility through reactions, side profiles, wide shots, inserts, and voice-over;
**Hybrid:** keep good native hero lines, and unify narration, phone voices, ambience, music, and problem audio in post.

All modes must be multitrack. Across shots in a continuous scene, maintain room tone/rain, use J/L cuts, lower BGM to make room for dialogue, and avoid restarting sound every time the picture cuts.

For multi-person native dialogue, record for each line speaker_id, visible/off-screen source, the lip-sync subject in the shot, non-speaker mouth state, and turn order. When staff explain rules, the staff member must visibly speak on screen, or explicitly use sourced broadcast/voice-over; other characters keep their mouths closed and only give listening reactions. Wrong person speaking, voice crossing characters, or non-speakers with synced lips is P0, and must not be explained away with subtitles.

Gated generation order

Validate the story and beat plan;
Validate character, style, world, and prop masters;
Generate the hook segment or the highest-risk segment first;
Download and directly inspect the actual output;
For a 30-second segment, make a 12-frame contact sheet and inspect at least the beginning, 25%, 50%, 75%, and end;
Check contact peaks, clipping, and physics in slow motion;
When audio is present, transcribe it and listen manually;
Use evidence to mark accepted、repair or rejected；
extract the acceptance tail frame and update the continuity state;
only then unlock dependent segments.

Can be parallelized: independent masters, music exploration, vocal lines, non-dependent inserts, and alternate hooks. Cannot be parallelized: continuous story segments whose accepted parent state has not yet been obtained.

Physics and anatomy gate

Read physics-and-failure-modes.md. Before generation, list the visible physical_contact_pointsand use action primitives that match the subject's anatomy. Human hands, animal claws, machines, liquids, cloth, vehicles, and magical effects use different contact rules.

Required order:接近 → 接触 → 施力 → 响应 → 稳定结果。

"reversing, turning, detaching, colliding" must be compiled into physical_start_state → control_input → world/screen motion vectors → phase_sequence → physical_end_state. One generated segment carries only one high-risk precision physics event; complex events are split into establishment, peak, and consequence shots. If the prompt intent is correct but the actual on-screen direction is wrong, it is still judged a failure.

Reject body/object penetration, extra or missing animals, impossible support, prop teleportation, phone orientation reversal, clothing structure changes, unexplained resets, or invisible causality.

QA, repair, and editing

Use qa-report.md and qa-and-repair.md, running respectively:

**Story:** a cold viewer can retell goal → obstacle → choice/action → result;
**Identity/art style:** characters, animals, products, clothing, materials, and medium remain consistent with the master;
**Physics/continuity:** contact, direction, state, and seams are credible;
**Sound:** dialogue, vocal lines, loudness, ambience, BGM, and bridging are continuous.

For comedy/animated performance, additionally do a silent acceptance check: the viewer can point out the joke's normal state, visible cause, information gap, reaction frame, reveal, and consequence for the main line; characters have at least three distinguishable performance states within a 15–30 second segment. If "where the thing went" or "why it is funny" can only be explained through dialogue, it is judged a story/performance failure.

Repair only the smallest failing unit: use editing for local defects, bridging/continuation for seam issues, regenerate only for identity/anatomy/causal collapse, and prioritize remixing for pure audio issues. Do not use music to cover up story failure.

For multiple candidates, first pass the MUST-HIT binary gate, then use candidate-scorecard.csv to compare story, identity, space, physics, performance, sound, and editability; when possible, blind key candidates to model, seed, cost, and generation order. Reviewers are responsible first for discovery and grading, and creators/repairers then decide how to fix, avoiding the same round defending its own results.

When entering the final cut stage, read editing-delivery-and-release.md, use edit-timeline.csv to include only accepted, genuinely usable intervals, unify the working format and delivery configuration, preserve subtitles/artistic text and audio stems, and check for missing media, timeline gaps, black/frozen/duplicate frames, loudness, subtitle safe areas, and horizontal/vertical versions. Generation complete does not equal editing complete, and upload success does not equal release approval.

Deliver according to task scope

Do not impose the full production package on a local request. Create only the files the next step will actually read or verify:

**Diagnosis/advice:** problem evidence, priority, and minimal repair; do not generate blank CSV or production directories.
**Story/screenplay:** production brief, story contract, script/beats; build corresponding masters only when identity, props, or season arcs truly have continuity.
**Storyboard/production package:** add coverage plan/frozen keyframes, generated segments, editing beats, reference authority, model capability cards, cost, and sound route.
**Execution:** Add request preview and fingerprint, idempotency key, task/recovery record, candidate lineage, cost ledger, real output, QA evidence, acceptance tail frame, and continuity status.
**Final cut/release:** Add the official timeline with source chain, media normalization, mixing, subtitles/artistic text, cover, encoding, and cold-audience report.

Reuse when existing assets or upstream state are sufficient; do not regenerate for the sake of "process completeness." Unified naming:BLOCK_ID_vN.mp4、_tail.png、_contact_12.jpg、_audio.wav、_QA.md。

Evidence boundary

Project documentation must distinguish:

**Documented capability:** Verify from the current official parameter page or interface structure;
**Measured result:** Observed under a specific model, entry point, date, and configuration;
**Creative preference:** Actively chosen for the current story;
**Inference:** Possible but not directly verified.

Dynamic prices, duration limits, or model quality must not be written as eternal facts.