Communitygithub.com

storyboard-sheet

把任意故事、脚本、短剧段落拆成镜头单,逐格生成统一风格的分镜画面,自动拼成带镜头号与时长标注的故事板(多格拼板图)。当用户要求“生成故事板/分镜板/故事板九宫格/分镜图/镜头板”,或需要把文字故事、影视段落、AI 视频脚本做成可视化分镜预览时使用。

storyboard-sheet 是什么?

storyboard-sheet is a Claude Code agent skill that 把任意故事、脚本、短剧段落拆成镜头单,逐格生成统一风格的分镜画面,自动拼成带镜头号与时长标注的故事板(多格拼板图)。当用户要求“生成故事板/分镜板/故事板九宫格/分镜图/镜头板”,或需要把文字故事、影视段落、AI 视频脚本做成可视化分镜预览时使用。.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/freedonplus/storyboard-sheet/tree/HEAD

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

Storyboard Sheet

Stability is not about improvisation, but about fixing this pipeline: Content Card → Shot List → Static Validation → Per-Frame Generation → Per-Frame Quality Check → Traceback Write-Back → Panel Composition with Overlaid Text → Final Board Review & Delivery. Every step's specification is locked in; only the story content changes, never the structure. Three gates enforce quality: Static validation must pass before image generation, per-frame QA must pass before panel composition, and original prompts must be written back before delivery.

Execution Flow

  1. Fill the content card. See prompt-templates.md for fields: film name & genre, character anchors, scenes, blocking card, spatial blueprint & camera position table, state card, lighting & texture, aspect ratio & frame count, shot list, and negative items. If anything is missing, the Agent fills it reasonably based on the story—do not ask the user for style preferences.
  2. Decompose the shot list. Split by content events, not by word count or visual appeal; one major action change per frame. Write it as a JSON isomorphic to assets/shots.example.json, strictly following field names and conventions. All reused sentences across frames are centralized in the top-level anchors block (single source of truth): style / characters / blueprint / blocking / axis / lighting / props / negative. Each frame's prompt is assembled from these eight original sentences—no frame may write its own variants. The camera positions table goes into cameras; each frame must have camera / in_state / out_state.
  3. Static validation (no validation, no image generation). First run scripts/validate-shots.ps1: check required fields, frame capacity, consecutive shot numbers, time ranges back-to-back, in_state/out_state word-for-word relay, whether each prompt contains all anchors as original sentences, whether camera references exist, and reference image chain direction. If any check FAILS, fix the JSON first—do not proceed to image generation.
  4. Per-frame storyboard image generation. A single image_gen call can submit all frames of the same scene (multiple items in request_list), each at 1536×864. Each frame's prompt follows the template from prompt-templates.md: global texture sentence + character anchor sentence + scene sentence + blocking anchor sentence + proportion anchor sentence + state card + entering state + this frame's action sentence + exiting state + negative sentence. All anchor sentences are copied verbatim across frames. Within the same scene, follow the camera position table order for the reference image chain: pass the previous frame's output as a reference image to the next frame, with the fixed instruction: "Only reference spatial layout, character proportions, ground relationships, and lighting direction; do not copy composition." When changing scenes, break the chain and start fresh with the first frame. See continuity.md for details.
  5. Per-frame quality control (no single-frame pass, no panel composition). As soon as each frame is produced, immediately check it against the ten-item list in prompt-templates.md: whether face type/clothing matches anchors, whether blocking is on the blueprint coordinates, whether left/right follows the axis rule, whether proportions match the anchor, whether lighting state/state card matches real semantics, whether light position and color temperature are consistent, whether the entering state connects from the previous frame, and whether there are any text watermarks. If unqualified, regenerate in place: Modify only the single faulty item in the prompt, not the anchor sentences. If the same frame is regenerated twice and still drifts, pass the first frame's output as a dual reference. If even that fails, downgrade the shot type (close-ups are more stable than wide shots).
  6. Traceback write-back (no write-back, no delivery). After each frame is finalized, write four items back into the corresponding shot in the shot list JSON: prompt (the complete original sentence submitted to the image generation model, verbatim), seed (if returned by the tool), ref (the reference image ID chain used for this frame, e.g., "SH003" or "SH001+SH003"), qa (quality check conclusions and notes). This JSON is the sole traceability archive: anyone who gets it can reproduce frame by frame and pinpoint which sentence caused a continuity error.
  7. Panel composition with overlaid text. Use scripts/build-board.ps1 (parameterized: spec JSON + frames directory + output path), following the proportion formula in output-spec.md to determine frame count and canvas. Shot numbers, durations, and shot names are overlaid as semi-transparent bars at the bottom of each frame during panel composition—do not let the image model draw any text. Before composing, run validate-shots.ps1 -PanelsDir <frames directory> -RequirePanels once more.
  8. Final board review and delivery. After composition, review the whole board: whether adjacent frames connect, whether character left/right positioning is consistent across the entire board, whether scene fixtures (stop lines, curbs, door frames, etc.) are consistent in form throughout, and whether title card information is correct. If issues are found, go back to step 5 and regenerate only the problematic frames. Deliver the composed board image (a PNG original + a JPG for easy viewing) along with the shot list JSON. Deliver individual frames only when the user requests them.

Hard Rules

  • Three Gates: Static validation of the shot list not passed → no image generation; per-frame quality check not passed → no panel composition; prompt/seed/ref/qa not written back → no delivery.
  • n×n grid + each frame 16:9 = whole board strict 16:9. 3×3 uses 1280×720 frames to compose 3840×2160; 4 frames use 2×2; 6 frames use 3×2 (extra wide) or 2×3 (portrait); 9 frames use 3×3. Non-square grid board ratio = (number of columns × 16) : (number of rows × 9).
  • Anchors as single source of truth: Characters, blocking, blueprint, lighting, state card, and negative items each have only one original sentence, living in the top-level anchors block. Each frame's prompt copies these original sentences verbatim. Do not write "same as above" or "maintain consistent styling"—do not rewrite near-synonyms per frame. A character's face changing across frames is the most common failure in panel composition.
  • Blocking uses scene fixtures as reference, copied verbatim per frame: stop lines, the Nth white line, curb edges, table edges, door frames, guardrail posts. Do not use "foreground/background/left side of the frame" to describe character positions—these change with shot type and camera position and are the number one cause of blocking drift. Displacement is only written as direction and distance relative to the anchor ("take one step forward, cross the stop line").
  • Scene consistency trio: Blueprint (spatial coordinates) + state card (semantics of mutable objects) + scene fixture forms (material/color/quantity)—all copied verbatim per frame across the group. Changing scenes is only allowed when specified; mark scene_break: true in the shot list and break the reference image chain.
  • Lighting position sentence is uniform across the whole group; it is only modified if the narrative actually requires a time change, and the change must be noted in the shot list.
  • One action per frame: Starting state → path/contact point → visible result. Multi-step actions are split into adjacent frames.
  • Adjacent frames must relay: Each frame's out_state must match the next frame's in_state word for word (position/orientation/action phase/contact point/expression intensity—five items). Do not use "basically the same." Scene breaks use scene_break: true explicitly—no silent disconnects.
  • Don't cross the axis; move according to the camera position table: All cameras are on the same side; characters do not swap left and right on screen. Adjacent camera angles ≤ 90°, height difference ≤ one body height, and shot types do not jump two levels.
  • Write the proportion anchor into every frame: Provide a scene fixture's fixed size (white line 45 cm, curb 15 cm, table width 60 cm). Character measurements are derived from this anchor to prevent sudden size changes.
  • State card copied verbatim per frame; lighting states must match real semantics: Pedestrian red light = standing red human figure, pedestrian green light = walking green human figure; vehicle light = plain colored disc. Specify which light faces whom, and ensure traffic flow matches right-of-way. Do not mix human figures inside discs, green numbers with red lights, or other hybrids.
  • Cut points must state what is being connected: No "natural transition" or "smooth cut."
  • Replace vague words with concrete specifics: "Premium feel," "stunning," "cinematic," "atmospheric"—all are replaced with specific details about lighting, material, speed, distance, and shot type.
  • Negative items always include: No text, subtitles, watermarks, numbers, or borders. Add more based on the genre (e.g., no blood spray, no modern firearms, no cartoonification).
  • If fewer frames than columns × rows and title_cell is enabled, the last cell is the title card; otherwise, leave the empty cell with a dark background—do not insert unrelated frames to fill.

Resources

FileResponsibility
references/prompt-templates.mdContent card, per-frame prompt template, full-board grid prompt, action and camera move writing, per-frame quality control gate, failure checklist
references/continuity.mdContinuity card: spatial blueprint, axis rules, camera position table, entry/exit state relay, reference image chain, proportion anchor, traceback write-back, self-check checklist
references/output-spec.mdFrame count and proportion formulas, resolution, overlaid text layout, shot list JSON field specification (including anchors and traceback fields)
assets/shots.example.jsonComplete shot list JSON example (8 frames + title card, including anchors/cameras/traceback fields)
assets/demo.jpgExample output: full board preview from the above shot list (8 frames + title card, 3×3, long side 2560 JPG)
scripts/validate-shots.ps1Shot list static validation: fields, relay, anchor phrase inclusion, camera reference, frame file existence
scripts/build-board.ps1Panel composition with overlaid text: reads JSON, outputs full board PNG and JPG preview

Both scripts depend on Windows built-in PowerShell and System.Drawing; no additional libraries are needed. Run commands:

# Before image generation: static validation (no pass, no generation)
& "<skill-dir>\scripts\validate-shots.ps1" -SpecPath <shots.json>

# Before panel composition: validate including frame files (fails on missing frames)
& "<skill-dir>\scripts\validate-shots.ps1" -SpecPath <shots.json> -PanelsDir <frames directory> -RequirePanels

# Panel composition with overlaid text
& "<skill-dir>\scripts\build-board.ps1" -SpecPath <shots.json> -PanelsDir <frames directory> -OutputPath <fullboard.png> -PreviewPath <preview.jpg>

The composition script crops and fills each cell by center (cover), not stretching the image; if frame dimensions differ from the cell aspect ratio, there will be no distortion.

See the example output at assets/demo.jpg: a full board preview of 8 frames + title card, corresponding to the shot list at assets/shots.example.json. Running validate-shots.ps1 first and then using the same JSON will reproduce the same structure.

相关技能