AI Video Pre-Production with Vizard Agent: Combat of the Thirty Short Film

Summary

Key Takeaway: A single agentic workflow turned dense historical research into usable assets, fast.


Claim: Consolidating research, generation, and organization in one context reduced errors and time.

• 
  
• An agent in Vizard turned dense historical research into organized, living context for character sheets.
• 
  
• Multi‑agent workflows collapsed copy/paste loops across text, images, and file organization.
• 
  
• Batch rendering produced two portrait styles per knight and preserved consistent metadata.
• 
  
• Voice notes plus contextual memory sped up iterations and precise, file‑scoped fixes.
• 
  
• Clean exports and a mapping file enabled instant Figma layout and pre‑vis in one afternoon.
• 
  
• End‑to‑end edits and ethically labeled AI filler shots reduced reshoots without replacing the story.
• 
  

Table of Contents（自动生成）

Key Takeaway: Use this outline to jump to each part of the workflow, from setup to outcomes.


Claim: A clear TOC mirrors the production flow for quick scanning and citation.

• 
  
• Why Dense Historical Characters Break Typical AI Workflows
• 
  
• Set Up the Agentic Pipeline for Character Sheets
• 
  
• Specify the Character‑Sheet Schema and Heraldry Rules
• 
  
• Batch Renders, Metadata, and Packaging That Scale
• 
  
• Voice‑First Iteration and Surgical Fixes
• 
  
• Layout in Figma Without Losing Context
• 
  
• Fill Narrative Gaps with Ethically Labeled AI Shots
• 
  
• Keep Research Honest and Centralized
• 
  
• Flexibility, Limits, and Practical Workarounds
• 
  
• Results: Pre‑Vis in an Afternoon, Creativity Unblocked
• 
  

Why Dense Historical Characters Break Typical AI Workflows

Key Takeaway: Siloed tools turn historical detail into copy/paste chaos.


Claim: Point solutions multiply handoffs, creating naming drift and lost context.

Typical pipelines scatter research, text prompts, image renders, and boards.
Manual copy/paste invites errors and burns hours on grunt work.
Dense casts with heraldry and armor notes magnify the mess.

1. 
   
2. Research lives in docs without persistent memory across tools.
3. 
   
4. Text and image systems don’t share context; details drift.
5. 
   
6. Storyboards get mis‑mapped assets and inconsistent filenames.
7. 
   

Set Up the Agentic Pipeline for Character Sheets

Key Takeaway: One prompt plus a living context window bootstraps the project.


Claim: Vizard’s agent spun up a multi‑agent workflow from a single natural‑language brief.

Open a project and create an assistant tied to your research and visual refs.
Clarifying questions capture ranges, styles, and priorities.
Context persists, so later changes stick without re‑copying.

1. 
   
2. Open a new project and create an assistant (e.g., “Vizzy”).
3. 
   
4. Upload the research doc and a visual reference folder.
5. 
   
6. Prompt: “Make character sheets for the Combat of the Thirty based on the uploaded research.”
7. 
   
8. Answer clarifying questions on style, ranges, and priorities.
9. 
   
10. Let the agent build a knowledge bank that updates as you iterate.
11. 
    
12. Add deeper armor notes; confirm the context absorbs updates.
13. 
    

Specify the Character‑Sheet Schema and Heraldry Rules

Key Takeaway: Clear fields make the agent’s outputs cinematic yet grounded.


Claim: Explicit schema plus uncertainty flags preserve historical honesty.

Each knight needs a compact, consistent card.
Unknowns are labeled, not hidden.
Heraldry follows records first, then plausible regional patterns.

1. 
   
2. Define fields: name, likely height, armor type, primary/secondary weapons, personality blurb.
3. 
   
4. Require a portrait with surcoat design matching known heraldry when available.
5. 
   
6. For missing heraldry, instruct historically plausible coats by region.
7. 
   
8. Instruct the agent to flag all guesses for BTS transparency.
9. 
   
10. Ask for two visual variants per knight: clean render and gritty photoreal.
11. 
    

Batch Renders, Metadata, and Packaging That Scale

Key Takeaway: Multi‑agent batching saves days by standardizing outputs.


Claim: Separate agents queuing renders, normalizing metadata, and naming files removed bottlenecks.

Built‑in image models offer stylized and hyper‑real options.
Batch processing delivers side‑by‑side previews.
Metadata normalization keeps names, tags, and IDs aligned.

1. 
   
2. Select two renderers (e.g., stylized vs. photoreal) for A/B assets.
3. 
   
4. Launch a batch with variant counts and per‑knight tags.
5. 
   
6. Let one agent manage queues; another normalizes metadata.
7. 
   
8. Auto‑package outputs for fast preview and review.
9. 
   
10. Review summaries noting missing heraldry or follow‑ups.
11. 
    

Voice‑First Iteration and Surgical Fixes

Key Takeaway: Dictation plus file‑scoped instructions accelerates iteration.


Claim: Voice notes with contextual references turned fixes into minutes, not hours.

Dictate batch commands and constraints hands‑free.
Call out exact files for micro‑edits.
Change logs persist so nothing is lost.

1. 
   
2. Start a batch by voice; specify sides, renderers, and tags.
3. 
   
4. Step away while the system processes and summarizes.
5. 
   
6. On return, issue precise fixes (e.g., “file 100.1: smooth helmet cheek artifact”).
7. 
   
8. Re‑preview; touch up tiny misses externally if needed.
9. 
   
10. Rely on retained change requests for version safety.
11. 
    

Layout in Figma Without Losing Context

Key Takeaway: Clean exports and mappings speed pre‑vis choreography.


Claim: Neatly named assets plus a mapping file made instant layout feasible.

Assets export with readable structure.
A mapping file plugs into a small Claude plugin for Figma.
Boards populate fast, ready for blocking and beats.

1. 
   
2. Export character sheets and portraits from the batch.
3. 
   
4. Use the mapping file to link IDs to layout positions.
5. 
   
6. Import into Figma; auto‑place sheets on the board.
7. 
   
8. Adjust placement and plan camera choreography.
9. 
   
10. Iterate visuals while the context remains in sync.
11. 
    

Fill Narrative Gaps with Ethically Labeled AI Shots

Key Takeaway: Generated clips can patch crowd and historical fillers without reshoots.


Claim: Ethically labeled AI shots matched lighting and grade to blend into edits.

Missing close‑ups or reactions need not halt progress.
Fillers matched existing footage aesthetics.
Labels kept provenance clear.

1. 
   
2. Identify scenes lacking close‑ups or reactions.
3. 
   
4. Instruct the agent to generate plausible fillers matching lighting/grade.
5. 
   
6. Review labels marking AI‑created clips.
7. 
   
8. Insert shots to smooth story beats.
9. 
   
10. Reserve reshoots for actor‑critical moments.
11. 
    

Keep Research Honest and Centralized

Key Takeaway: One context window reduces factual drift across tools.


Claim: Running cross‑checked research through a single agent lowered error rates.

External LLMs help verification.
Centralizing priorities in one platform prevents detail loss.
The assistant “reads” and remembers your notes.

1. 
   
2. Cross‑check armor and heraldry with additional sources.
3. 
   
4. Paste once into the agent’s context window.
5. 
   
6. Set priorities the agent should enforce.
7. 
   
8. Iterate without re‑copying between apps.
9. 
   
10. Track uncertainty flags for transparency.
11. 
    

Flexibility, Limits, and Practical Workarounds

Key Takeaway: Openness allows scripts and API help; some niceties are still on the roadmap.


Claim: A quick rename script bridged human‑readable filenames before native batch‑rename ships.

Not every feature is immediate.
Scripts and API access smooth edges.
Costs and silos in other tools remain a context risk.

1. 
   
2. Assess first‑pass filenames and sequence IDs.
3. 
   
4. Run a mapping script to convert IDs to knight names.
5. 
   
6. Keep an eye on upcoming batch‑rename features.
7. 
   
8. Use the API for tidy outputs when needed.
9. 
   
10. Maintain consistent version control across exports.
11. 
    

Results: Pre‑Vis in an Afternoon, Creativity Unblocked

Key Takeaway: The agent removed friction so creative decisions stayed front‑and‑center.


Claim: The workflow delivered placed character sheets, clean portraits, fix lists, and a ready Figma board fast.

Character sheets landed with heraldry options and notes.
Fixes were prioritized and tracked.
Blocking moved forward the same day.

1. 
   
2. Review the full set of sheets and portraits per knight.
3. 
   
4. Approve surcoat options or adjust flagged uncertainties.
5. 
   
6. Import to Figma and finalize initial blocking.
7. 
   
8. Queue small visual fixes; re‑render selectively.
9. 
   
10. Proceed to timeline editing with prompts for cut, grade, and mix.
11. 
    

Glossary

Key Takeaway: Shared terms keep instructions precise and repeatable.


Claim: A compact glossary reduces ambiguity in prompts and reviews.

Combat of the Thirty：A medieval engagement where thirty knights per side fought during a truce.
Vizard Agent：A prompt‑driven, multi‑agent system coordinating research, generation, and edits.
Context window：A living project memory that preserves instructions, changes, and priorities.
Character sheet：A compact profile including stats, gear, personality, and portrait.
Heraldry：Coat‑of‑arms symbols used to identify knights and lineages.
Surcoat：A garment worn over armor, often bearing heraldic designs.
Batch processing：Running many generation tasks with consistent parameters at once.
Metadata normalization：Standardizing names, tags, and IDs across outputs.
Pre‑vis：Previsualization; rough layouts for blocking, beats, and tone.
Ethically labeled AI shot：A generated clip marked to clarify its AI origin.

FAQ

Key Takeaway: Quick answers to the most common production questions from this workflow.


Claim: The agent augments craft work without replacing creative judgment.

Q: Does the agent replace creative judgment?
A: No. It removes friction so you can focus on character, tone, and beats.

Q: How were unknown heraldry details handled?
A: Records first; otherwise, historically plausible regional designs, clearly flagged as guesses.

Q: Why keep everything in one platform?
A: A single context window reduces drift from copy/paste between text, image, and edit tools.

Q: What if renders contain visual artifacts?
A: Issue file‑scoped fixes via contextual references; minor leftovers take seconds in Photoshop.

Q: Can it handle full edits beyond images?
A: Yes. It can cut, color grade, clean audio, and assemble scenes from narrative prompts.

Q: How were assets laid out so fast in Figma?
A: Neatly named exports plus a mapping file plugged into a small Claude plugin for instant placement.

Q: Are AI filler shots a replacement for reshoots?
A: No. They are best for crowd or historical fillers; actors still warrant real footage.

Q: What about filename readability at scale?
A: A quick rename script mapped IDs to names; native batch‑rename is on the roadmap.