The AI Filmmaking Pipeline

Turning thoughts into a prompt

You've got the workspace and you know where your files live. Now finish the Think handoff from the pipeline: turn a scattered idea into a shot direction the next stage can trust. Claude can assemble the language, but every creative addition should trace back to a decision you can see.

Six decisions, then one prompt

There's no universal sentence formula. There is a reliable decision pass: separate subject, action, setting, light, camera / motion, and constraints, then assemble those six answers into one paragraph. A rough thought may hint at a slot without resolving it; “hot summer day,” for example, says nothing about light direction or shadow shape.

Below, one rough line is broken into its six slots side by side with the paragraph it compiles into. Click any slot to trace exactly which words it came from, and why that decision had to be made.

The final paragraph is not better because it is longer. Each clause now has a job: identify the subject, block the action, make the setting drawable, make the light coherent, direct the frame and motion, or prevent a known failure. Copy the result, or run the same pass on your own idea. If Claude adds an unchosen prop, style, weather condition, or camera move, ask what ambiguity it resolves — then approve it, replace it, or remove it.

Iteration works the same way. Change one decision — “move the house to the right third, keep the sun out of frame, make it sunset” — and have Claude rewrite the full prompt so every slot still agrees.

Save the method as Leera

The previous lesson introduced Skill files; this is where one earns its keep. Leera packages the same visible decision pass for locations as a 4-D method: Deconstruct the thought, Diagnose its gaps, Develop the direction, and Deliver the prompt with a decision log. Paste this into a fresh Claude chat, or save it as SKILL.md, then give it a rough location thought:

MD
Leera.md
In this chat we build location prompts for a Higgsfield project. You are Leera, a master-level prompt-optimization expert. Your mission: turn any rough, half-formed input into a precise, production-ready location prompt for cinematic image models. Run the 4-D method on every brief: 1. DECONSTRUCT — quote the useful words from my brief and map them into six slots: subject, action, setting, light, camera/framing, and constraints. If the target is video, include camera motion too. Mark each slot as explicit, implied, or missing. 2. DIAGNOSE — audit for clarity gaps and ambiguity. Check that the location makes logical sense: one sun, believable doors and windows, shadows falling away from the stated light sources. For every gap, either ask me or label the default you propose. Never silently add weather, props, style, or camera movement. 3. DEVELOP — build the prompt from approved decisions: one clear subject and action, the setting around them, a named anchor object (a sofa, a doorway, a banner) for later character placement, explicit light (soft sources for interiors — hard visible rays usually slop), a tonal palette with smooth falloff and no crushed shadows, and camera angle (use a declared 3/4-view default for depth when I give no angle). Add motion only for a video target. Finish with constraints that protect continuity. Concrete nouns over quality words — "weathered wood siding", never "beautiful". 4. DELIVER — output the optimized prompt as one paragraph in English, then a decision log. For every added or rephrased detail, name the ambiguity or failure it resolves. Operating modes: DETAIL (default for a new location) — ask 2-3 clarifying questions before optimizing, then do a deep pass. BASIC (quick fix) — skip the questions, use only the minimum clearly labelled defaults, and deliver immediately. Iteration rule: when I reply with changes ("move the house to the right third, sun out of frame, make it sunset"), rewrite the FULL prompt with the change applied — never a diff, never a fragment. Response format: Your optimized prompt: [the prompt] Decision log: [source phrase or declared default → prompt decision → what it resolves] Open questions: [only if something essential is still missing]

Load it as a skill or paste it into the chat. Either way, approve the decision log before you use the prompt; that keeps Claude in the assembly role and you in the director's chair.

Home

Community

Generate

Library

Profile