kaomei/stickman-video-director / skills/directing-stickman-videos
directing-stickman-videos
UnverifiedTurn topics, requirements, articles or scripts into stickman (stick figure) animated videos: explainers, short stories and faceless YouTube, Shorts or TikTok content, in styles from minimal line art and retro TV to pixel art, rubber hose and meme animation. Model Engine (A) delivers storyboards and standalone prompts, or uses available authorized tools to generate and finish videos; Code Engine (B, formerly Autopilot) produces narrated, captioned illustration-and-micro-animation videos through configured generation tools and a local renderer. Also use for 火柴人视频 requests, and to guide installation handoff, first use and setup of this specific Skill.
Directing Stickman Videos
Two modes, one Skill
Use A · Model Engine(模型引擎) as the canonical A mode name in Chinese guidance. Recognize “导演模式 / Director” as its previous name. Likewise use B · Code Engine(代码引擎) as the canonical B mode name, and recognize “Autopilot / 自动成片” as its previous name; existing autopilot file names and commands are unchanged. Keep the Skill name, A style IDs, existing paths and each style's visual and sound identity; the unified A workflow replaces legacy A delivery and fixed-duration rules.
A follows one staged workflow: confirm story/style/ratio/duration/sound and the storyboard → deliver standalone prompts with reference/settings guidance → offer optional built-in-browser generation of representative tests at the chosen model's lowest available resolution → after test acceptance, confirm final resolution and total credit budget → continuously produce, repair and finish the film. Prompt delivery is a complete stopping point. Reuse prior authorizations; an explicitly approved full production need not repeat a trial. Default browser-assisted generation to Google Flow for proximity to the showcased results, honoring the user's chosen platform. Read the unified A workflow; do not ask users to choose two routes upfront. Do not imply A always requires manual editing or that every installation can automatically operate a video service. Preserve the selected style's drawing, prompt/reference and actual-output acceptance requirements; use the shared A stages for delivery and execution authorization.
For every A1–A7 project, read the A production and review contract and independent audio timeline and listening contract, with the selected voice continuity profile. Distinguish generated source clips from edited shots, preserve the requested final runtime, fix the selected character/style/reference rules while varying the story, save execution progress and limit ineffective retries. Keep music, narration, native dialogue, effects and ambience on their appropriate final timelines; review actual acting, camera, continuity and sound. Repair the affected scope and reuse explicitly user-confirmed samples without repeating their review. Apply action and camera continuity at assembled joins, including each insert's entry and return.
Keep the name and known choices; clarify ambiguity once. Before planning, read mode execution boundaries, then the selected route:
-
A · Model Engine(模型引擎) / A1–A5: confirmed proposals and standalone video prompts, followed by optional authorized testing and production. A1 maps to Style 1 light/dark, A2 to Style 2/2A/2B, A3–A5 to Styles 3–5. Read A1–A5 direction and the shared unified A workflow before planning; for A3, A4 and A5 also read their Prehistoric, Historical Cartoon or Cozy Storytelling style reference.
-
A6 · Amber Noir: read A6 when selected, before planning; bypass the A1–A5 direction reference while following the shared A workflow. It owns reference-guided execution, Gemini Puck narration and finishing.
-
A7 · 橡皮管动画风 / Rubber Hose Animation: read production, four-block frozen prompts, visual references and actual-output acceptance. Default to the successful MV route: fixed character/style blocks + existing kitchen reference material + compact new scene and timed acting/camera → direct video generation. Use the image as an actual identity/drawing reference, not a compulsory kitchen first frame; no mandatory new still or low-resolution probe for every location. Reuse the two accepted expanded examples, preserving the thin-line two-tuft hero and 1930s ink-and-paint scenery. Review actual motion, contacts, drawing and sound; matching prompts/reference inputs do not certify output. A7 owns its Model Engine workflow and optional narration/captions; sample lyrics are not default requirements. Bypass A1–A5 rules and B renderers.
-
B · Code Engine(代码引擎): read Code Engine (Autopilot), then B1 Minimal Line Art, B2 Colorful Hand-Drawn, or B3 Retro TV. B3 also requires semantic lines and purposeful motion: plan cue-linked events and object-bound connectors automatically, then independently review every event/graphic in the actual MP4; geometry and pixel change do not establish meaning. B1/B2/B3 require sample-effect acceptance: capable agents target the samples; genuine tool limits permit disclosed best effort, never an equivalence claim.
-
B4 · Survival Thriller: read visuals, asset boundary, picture animation, audio, opening, typography and QA. Create story-specific artwork; references/showcases are quality standards only. Require independent
story_asset_originplus all existing reviews. Automatically plan and render story-led local picture changes through the body as well as animated text; no reminder to animate is needed. Preserve the white-glove drawing lock; limb animation is not a completion requirement. Apply the first-three-second visual escalation gate and story-led horror visibility gate, keep every opening/body/ending shot at most three seconds, and time meaningful action within that limit against events/actual speech. Crops, subtitle updates, micro-zooms and decoration do not reset an unchanged principal-picture state’s three-second dwell. Useb4-hook-v1for landscape and the approved portrait poster treatment for portrait through the helper. Every language/ratio needs separate independent actual-MP4 picture and text reviews; technical checks are not acceptance. -
B5 · Avatar Explainer: read B5 and the default first-preview workflow. Before composing a first preview or new story, read the latest approved filled Mei example, watch its exact approved film, and read the actual relay stage and Mei rig. Read the original Mei prompt and fill the Mei scene template. Apply directing structure and the motion contract: successive lead visuals and clear handoffs, source proportions, scene richness, and motivated physical action. CrazyMei + Spider remain the drawing/action standard in both native ratios; Jensen is historical only. Let the current words choose the ending; never force a side-by-side recap or summary to fill time. Draw the needed objects, states and contact poses before tweening, reusing approved identity/audio and suitable artwork. Use the installed executable artwork and relay kit, copied by default into each new project: source-drawn contours, complete figure/contact poses, whole cartoon animal states, safe composition and evaluated endpoint handoffs. Prefer graphic substitution and whole-object comic motion over unnecessary animal-joint simulation. Author, render and encode an internal candidate, then inspect that actual candidate once at normal speed with sound, checking motion relay, figures/contact, composition and semantic captions. Inspect suspect frames only as needed; repair and recheck affected intervals and neighbors. Default previews do not require independent storyboard rehearsal, plan-lock, blind PNG/seal/contextual rounds or per-event evidence clips. The strict formal protocol remains opt-in. State actual viewing/listening limitations without inventing a pass or replacing unavailable playback with repeated whole-image audits; user acceptance is never inferred.
-
B6 · 动感 Meme 风: read B6, the mandatory installed production workflow and character-binding contract. Turn the brief into a finished new film: assets, native shots, binding, sound and QA; a copied gallery or V90 rerender is insufficient. Default init provides compositor, puppet and spatial helpers without exposing the old V90 performance to new-story scenes. Bundled reference/test/example media are compare/study-only, even renamed; paint new story assets. Legacy tests remain diagnostic regardless of fixture/review edits. Explicit unchanged V90 reproduction uses its recipe, labelled
example-reproduction. Native 16:9/9:16 requires separately authored plans; the V90 performance currently validates 16:9 only. For new cast/actions use the installed executable character packages and compositor, not numeric pins alone; default projects require their final visible-alpha evidence. Reuse identity/view bindings across scene branches and background cast; verify complete chains, fixed lengths, elbow direction, occlusion, handoffs and grounded legs before extending. For an era/place, apply period and trend direction to clothing, hair, accessories and scenery while preserving the brushwork.
B6 authoring defaults to the faster iteration path: check character imports before new paid narration/full encoding, reuse compatible action timing helpers, then inspect a true lower-resolution 30fps draft through scene.draft_frame and context.viewport. Reuse accepted audio; draft/worklist results do not inherit review passes or replace final acceptance.
B6 works best in portrait (9:16). Sample 1's portrait Eiffel V44 is the strict primary drawing and animation standard, including its heavier relative limb stroke; Sample 2 V90 is secondary binding/engineering evidence. Archivo Black is B6’s user-approved default caption font, retaining the established size, placement, outline and phrase switching. Apply the Sample 1 acceptance matrix and motion execution and review gate: lock the primary reference, compare limbs at matched head width, check assets before promising a shot, inspect actual decoded playback with its implementation, and separate technical checks, creative review and user approval. New-story init includes the exact portrait reference, its style standard and the portable spatial helper. A plan label, finer V90 limbs or successful encode does not establish matched quality. All 35 detailed subchecks are mandatory review coverage, including the approved Archivo Black portrait caption profile/placement/raster, semantic timing, head/body paint, all-cast limbs, soft-light/material hierarchy, independent background life, purposeful camera and original-speed sound. An unchanged/synced subtitle is not evidence of matching the approved caption standard. A broad pass cannot hide an unchecked subcheck. Animation performance and spatial richness are mandatory acceptance items. On pending/needs-repair findings, continue authorized local correction, re-rendering and actual review; do not treat a failed test preview as completion or hand the unfinished acceptance work to the user. Preserve and link each rejected version’s repair findings.
- B7 · Pixel Theater: read B7, inspect both bundled approved full audiovisual references, then use reviewed project-local code with the bundled pixel/ending helpers. For 9:16, also read the native portrait and short-caption contract and inspect its approved internal 10-second reference. Apply the B7 evidence and actual-film review gate. B7 is not a shared-manifest style or an image-generation fallback; preserve pixel geometry, causal contact/action, semantic captions and story-led sound.
A and B share installation, not production contracts. The A1–A5 direction applies only to A1–A5; A6/A7 keep dedicated style references. The unified A workflow applies to A1–A7, never to B. Each B owns its visuals, captions, voice, pacing and SFX; no B1 fallback or A setup/clip/approval/text rules. Preserve unrelated styles.
B1–B6: read the installed workflow and run its style/ratio query before production.
B accepts topics, briefs, articles and scripts; default English unless overridden. Follow selected-provider execution: a missing adapter is not a provider prohibition. Use native named TTS without style descriptions; B4 allows secondary-role acting, not narrator tuning. Read the story-led sound system. Never automatically revoice accepted media.
Captions: B1–B4 and B6 single-line; B5 up to three semantic lines. Apply fixed style typography and actual local motion/audio checks in the common gate. B1/B2/B3/B5 voice credits are illustrative; Brian remains B4-only. New B5 real/complex presenters require an available image tool or approved supplied raster, not code-face substitutes; follow the production gate.
All B styles: native 16:9 and 9:16, under the dual-format gate. For B1–B3 portrait, also read the native portrait contract: ratio-specific readable captions, authored per-shot focal regions, safe-area/visible-action checks and independent 1× phone-size review are required. Use the approved internal 10-second B1/B2/B3 portrait samples as effect targets alongside the landscape references; B1 retains its handwritten hierarchy and sequential element builds, while B2/B3 retain their original visual languages. They are not additional README showcases. Their exact-clip user approval, technical diagnostics and each new output's independent review remain separate; an engineering render does not establish portrait acceptance.
First use and next steps
Read first-use and stage guidance after verified installation or on first invocation. Use the conversation/requested guidance language independently of video language or README edition. Reuse choices; do not replay welcome or add approvals. Follow A5 proactive character guidance and Optional A character references there without changing voice rules.
Read optional finishing after video approval or a packaging request; offer once, never block delivery. See that reference for the optional project credit and when to include a closing GitHub Star invitation.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |