Step5 Shots Prompts

• Shot prompt files under projects/<project-id>/prompts/ organized by platform:
  • prompts/kling/SHOT_XX_option_A.txt through option_D.txt
  • prompts/seedream/SHOT_XX_option_A.txt through option_C.txt
  • prompts/nanobanana/SHOT_XX_option_A.txt (static frame prompts)
  • prompts/suno/music_prompt.txt (music generation prompt)
• projects/<project-id>/prompts_index.json — compiled index of all prompts

Execute

1. Resolve project ID. Default to default if not specified.
2. Load all canon data:
   • Characters from bible/characters.json
   • Locations from bible/locations.json
   • Visual style from bible/visual_style.json
   • Cinematography from bible/cinematography.json
   • Shot list from bible/shot_list.json
3. For each shot in the shot list, generate prompts following the Prompt Construction Algorithm below.
4. Generate 3-4 variations per shot (A/B/C minimum, D optional).
5. Write prompt files to prompts/<platform>/.
6. Regenerate the prompt index:

npm run index -- --project <project-id>

7. Verify each shot in shot_list.json appears in prompts_index.json.

API Path (alternative)

If the server is running, generate prompts via the agent endpoint:

• POST /api/agents/prompt-runs?project=<id> — triggers prompt compilation for all shots

Prompt Construction Algorithm

For each shot, follow these steps:

Step A: Extract Shot Context

shot = shot_list.shots[i]
characters = shot.characters.map(c => lookup in characters.json)
location = lookup shot.location.id in locations.json
timing = shot.timing
intent = shot.intent
cameraIntent = shot.cameraIntent

Step B: Build Character Identity Anchors

For each character in the shot:

Primary characters (full anchor):

[physicalCore.age] [physicalCore.build], [physicalCore.skinTone].
[faceSignature.structure], [faceSignature.eyes], [faceSignature.hair].
[costume.default.description] with [costume.default.signature].

Secondary characters (compact anchor):

[physicalCore.age] [physicalCore.build], [faceSignature.hair], [costume.default.signature].

Step C: Build Location Anchor

[setting.type], [setting.scale], [setting.architecture].
[atmosphere.lighting]. [atmosphere.weather].
Colors: [atmosphere.colorPalette].
Visual anchors: [ALL items from visualAnchors array].

Step D: Build Prompt Body (platform-specific)

See Platform-Specific Best Practices below for each platform's structure.

Step E: Apply Negative Prompt

[visual_style.negativePromptBase] + [platform-specific additions]

Step F: Apply Variation Strategy

What NEVER varies across A/B/C/D:

• Character identity anchors (physical, face, costume)
• Location visual anchors and color palette
• Action/intent
• Visual style and forbidden elements
• Negative prompt

Step G: Write Prompt File

Each file follows this header format:

=== SHOT {ID} - Variation {VAR} ({LABEL}) ===
Shot: {ID} | Section: {SECTION} | Time: {START}s-{END}s ({DURATION}s)
Version: 2026-02-08
Variation: {VAR}

--- {PLATFORM} PROMPT ---
{PROMPT BODY}

--- NEGATIVE PROMPT ---
{NEGATIVE PROMPT}

--- DIRECTOR NOTES ---
{intent.why}. Emotional beat: {intent.emotionalBeat}.

Platform-Specific Best Practices

Kling 3.0 (Video)

• Max length: 500 characters for prompt body
• File pattern: prompts/kling/SHOT_XX_option_[A-D].txt
• Focus: Motion, camera movement, action transitions over 8 seconds

Prompt Structure:

1. Shot size + lens feel (e.g., "medium close-up, 35mm anamorphic")
2. Camera movement (e.g., "slow push in") or "static hold"
3. Character identity anchor (compact ~80 chars)
4. Action in one continuous take
5. Location anchor (compact ~60 chars)
6. Lighting and color treatment
7. Style tag: "Photorealistic cinematic style, natural lighting, coherent motion over 8s."
8. Negative prompt line

Forbidden in Kling prompts:

• Scene cuts or location changes
• Multiple actions ("she walks then sits")
• Camera movements from cinematography.json.cameraMovement.forbidden
• Motion verbs that imply teleportation

SeedDream 4.5 (Image Pairs)

• Max length: 2000 characters total
• File pattern: prompts/seedream/SHOT_XX_option_[A-C].txt
• Focus: Dual-frame state change

Prompt Structure ("Create 2 Images"):

Create 2 Images.

Image 1: [Complete scene setup]
[Character full identity anchor]
[Location with all visual anchors]
[Action state A — the starting moment]
[Camera: framing, lens, composition]
[Lighting and atmosphere]

Image 2: Same [character description], same location.
[What changed: emotional shift, subtle pose change, light change]
[Emotional beat: "emotional weight has shifted from X to Y"]
[Same camera framing maintained]

SeedDream Rules:

• Frame 1 = complete scene setup
• Frame 2 = same character/location + describe the shift
• Use "emotional weight has shifted" language, not forced expressions like "she smiles"
• All visual anchors from location must appear in both frames
• Character identity anchors must appear in both frames

Variation Strategy:

• A: Standard framing, full descriptions
• B: Intimate close-up ("Extreme close-up, shallow DOF"), trim location detail, push camera in
• C: Wide cinematic ("Ultra-wide cinematic, IMAX format"), full location, environment focus

Nano Banana Pro 3 (Static Images)

• File pattern: prompts/nanobanana/SHOT_XX_option_[A-D].txt
• Focus: Single static frame, zero motion

Required Elements:

• Static scene composition only
• Character identity anchors
• Location visual anchors
• Style and lighting descriptors
• Composition language (foreground/background/depth)
• Negative prompt including motion verbs

Motion Verb Conversion Table:

Suno (Music)

• File pattern: prompts/suno/music_prompt.txt
• Focus: Pure musical language — no visual terms

Structure: Genre → Mood arc → Tempo/BPM → Instrumentation → Production style → Vocal style → Song structure

Forbidden terms: camera, lens, shot, frame, light, shadow, neon, scene, character. See Step 2 skill for the complete Musical Translation Guide.

Quality Checks

Before saving prompts, verify for each file:

1. Prompt matches platform-specific format (correct sections, within character limit).
2. All required identity anchors are present (character + location).
3. No cross-shot references ("as seen in SHOT_01", "continuing from the previous shot").
4. Negative prompt is present and includes visual_style.json.negativePromptBase.
5. No forbidden camera movements appear in Kling prompts.
6. Nano Banana prompts contain zero motion verbs.
7. Suno prompts contain zero visual/camera language.
8. Variations A/B/C/D differ only in camera/framing, not in identity anchors or action.
9. All prompts include version/date metadata in the file header.

Completeness Check

project=default
node -e "
  const fs = require('fs');
  const base = 'projects/' + process.env.P + '/';
  let ok = true;
  // Check prompts_index.json
  const idxPath = base + 'prompts_index.json';
  if (!fs.existsSync(idxPath)) { console.error('Missing: prompts_index.json'); process.exit(1); }
  try {
    const idx = JSON.parse(fs.readFileSync(idxPath, 'utf8'));
    if (!Array.isArray(idx.shots) || idx.shots.length === 0) {
      console.error('No shots in prompts index'); ok = false;
    } else {
      console.log('Indexed shots: ' + idx.shots.length);
    }
  } catch (e) { console.error('Invalid index: ' + e.message); ok = false; }
  // Check shot_list coverage
  try {
    const shots = JSON.parse(fs.readFileSync(base + 'bible/shot_list.json', 'utf8'));
    const idx = JSON.parse(fs.readFileSync(idxPath, 'utf8'));
    const indexedIds = new Set((idx.shots || []).map(s => s.id || s.shotId));
    for (const shot of (shots.shots || [])) {
      if (!indexedIds.has(shot.id)) {
        console.error('Shot ' + shot.id + ' not in prompts index'); ok = false;
      }
    }
    if (ok) console.log('All shots indexed: OK');
  } catch (e) { console.error('Coverage check: ' + e.message); }
  // Check prompt directories exist
  const promptDir = base + 'prompts/';
  if (fs.existsSync(promptDir)) {
    const dirs = fs.readdirSync(promptDir).filter(f => fs.statSync(promptDir + f).isDirectory());
    console.log('Prompt directories: ' + dirs.join(', '));
  }
  if (!ok) process.exit(1);
  console.log('Step 5 complete.');
" P="$project"

LLM Guidance

• Prompt text must be standalone. Never reference "the previous shot" or "as established earlier." Each prompt must contain all visual information needed to render the shot in isolation.
• Output format: Raw prompt text with section headers as shown in the file format above. No markdown code fences around the content.
• The most common mistake is forgetting identity anchors. Every prompt must describe the character and location as if the reader has never seen them before.
• When generating many prompts, process in batches of 5-10 shots to maintain quality.
• After generating all prompts, always run npm run index -- --project <id> to rebuild the index.
• Reference skills/_shared/references/universal-prompt-rules.md for platform constraints and identity anchor methodology.
• Reference references/system-prompt-best-practices.md for detailed per-platform guidance.

Step 5 is complete only when the prompts index is populated and every shot has prompt files for the target platforms.