The bottleneck is no longer typing, coding, drawing, or rendering.
The bottleneck is your taste, your judgment, and your domain knowledge.
"It's the democratization of storytelling at a level that has never happened in the existence of humankind."
— Jeffrey Katzenberg, DreamWorks Animation
Models · Apps · Harnesses
Three layers to understand. The same model behaves completely differently depending on how it's harnessed.
🧠
Model
The brain. How smart it is.
Claude Opus · GPT-5.2 · Gemini 3 Pro
→
📱
App
The interface you talk to.
claude.ai · chatgpt.com · NotebookLM
→
⚡
Harness
Tools + rules that let it do work.
Claude Code · Codex · Cowork · Deep Research
The Writer: Claude Opus
Use for: Script breakdowns, character bibles, emotional logic. Harness: Claude Code, Cowork, Deep Research.
The Producer: Gemini 3 Pro
Use for: Video dailies, 1M-token context, production PDFs. Harness: NotebookLM, Deep Research, Antigravity.
The Executor: GPT-5.2
Use for: Rapid tool execution, dashboards, automation. Harness: Codex, Deep Research, Canvas.
The Sovereign: Llama 3 / DeepSeek
Use for: Zero API bills, total data privacy, Setswana fine-tuning. Harness: Run locally, your rules.
⚠️ Free tiers give you weaker models by default. Always manually select the frontier model.
The Operator Skill Stack
You are not prompting. You are managing.
🧩
Decomposition
Break a complex goal into micro-steps the AI cannot fail at. The skill is knowing where to cut.
📂
Workspace Design
Structure folders and write rules so agents find their own context. Curation over injection.
⚖️
Evaluation & Taste
The model always gives you an answer. Is it production-grade? Your domain expertise is the moat.
🔁
Iteration & Oversight
Know when to push the model harder, when to hand off, and when to just fix it yourself.
"An AI that does things is fundamentally more useful than an AI that says things."
— Prof. Ethan Mollick, Wharton
Workspace Design in Practice
A single file teaches the agent who you are and how you work. This is a real example for a UB educator.
CLAUDE.md — University of Botswana Media Studies
# Connie Rapoo — Global Agent Instructions## Who I Am
Senior Lecturer, Dept of Media Studies, University of Botswana.
Primary work:
- Supervising 8 student short-film productions per semester
- Research papers on Southern African cinema (LaTeX)
- Managing the UB–Denmark Film Konnections collaboration
- Coordinating visiting lecturer schedules
## Confirmation GuidelinesDO ask before:
- Sending emails to external partners or the Ministry
- Writing to shared Google Docs or Overleaf projects
- Any action involving student grades or records
DO NOT ask before:
- Routine file creation, reading, or organizing
- After I approve a plan — just execute it
- Retrying after errors — fix and re-run
When in doubt: bias toward action.
## Technology Stack| System | Stored | Access |
|--------------|--------------------------|------------|
| Google Drive | Course materials, briefs | MCP/local |
| Overleaf | Research papers (LaTeX) | GitHub CLI |
| UB Portal | Student records, grades | Browser |
| Local Dirs | Scripts, call sheets | Filesystem |
| Gmail | All communication | MCP |
## Communication & Style
- Concise and direct
- Formal tone for Ministry/Dean; warm for students
- I use voice dictation — extract the actionable request
- Default to Setswana terminology with English gloss
🎯 Identity — The agent reads this first. Your title, projects, domain. Replaces re-explaining yourself every session.
⚡ Permission Boundaries — The most-iterated section. "Bias toward action" is the key line — without it, the agent asks permission for everything.
🗺️ Tech Map — Without this table, the agent guesses where files live — and gets it wrong. This is its mental model of your digital life.
📂 Workspace — This IS the skill from the previous slide. Structure folders well → the agent finds its own context. Zero copy-paste.
🗣️ Cultural Encoding — Setswana-first ensures AI respects local norms. Your expertise becomes the agent's default behavior.
What is an "Agent"?
A reasoning model trapped in a loop, given tools to affect the real world.
🎯
1. Observe
Read your prompt. Scan the workspace. Understand the goal.
→
🧠
2. Plan
Chain-of-thought. Break task into steps. Pick which tools to use.
→
🔧
3. Act
Read & write files. Run code. Browse web. Call APIs.
→
🔍
4. Evaluate
Did it work? If not → loop back. If yes → deliver.
If Step 4 fails → loops back to Step 2 autonomously. No human needed.
💡
Example:"Read the Morero treatment, build a shot list, and export it as a printable PDF."
The agent reads the .docx → extracts scenes → generates the list → writes HTML → converts to PDF → catches a missing font → installs it → re-exports → delivers.
Chatbot: you type, it replies
Agent: you give a goal, it works until it's done
Agents in Action: Pre-Production
Watch the CLI Agent (pi) read, write, self-correct, and deliver.
pi — Autonomous CLI Agent
$
OUTPUT GENERATED:
📄 outline.md (8 narrative beats)
🎥 shotlist.md (24 setup descriptions)
📋 callsheet.md (Cast & Crew list)
🌐 dashboard.html (Interactive UI)
Notice the self-correction step. That is reasoning in practice.
The Prompt IS the Shot List
Each layer maps 1:1 to a film skill you already have.
Prompt Layer
Film Skill
What We Write
1. Subject & Action
Cast Direction
"A Batswana mother clutches a near-empty plastic water bottle, child gripping her skirt"
2. Context & Setting
Production Design
"Cracked laterite earth, military checkpoint with razor wire, government water sign"
3. Composition
Blocking
"Queue of hundreds recedes into heat haze, woman foreground right third"
4. Lighting
Gaffer Notes
"Harsh equatorial noon, bleached-out sky, dust particles in backlight"
Your film training IS prompt engineering.Let's prove it ↓
Same Scene. Three Prompt Levels.
SubjectContextCompositionLightingCamera & Style
A woman waiting in a water queue in Botswana.
Layer 1 only
A tired Motswana mother holding a plastic water bottlewaits in a long queue outside a government water depot in a drought-stricken village. Cracked earth,harsh midday sun.Other people wait silently behind her.
+ Context, Composition, Lighting
A weathered Motswana mother in her 40s clutches a scratched plastic water bottle against her chest,standing third in a silent queue outside a concrete government water depot with a faded blue government logo on its metal gate. Cracked laterite earth extends in every direction with zero vegetation.Behind her, dozens of people stretch into the heat haze, each holding a single container.Her headwrap is sun-bleached indigo. Sweat beads on her forehead.Harsh overhead equatorial sun at noon creates deep eye-socket shadows.Dystopian drama, shot on ARRI Alexa 65 with 40 mm anamorphic lens, shallow depth of field isolating her face, desaturated warm tones, Denis Villeneuve color palette, 2.39:1 aspect ratio.
All 5 Layers = Cinema
The New VFX Pipeline
Nano Banana 2 Edit: describe what to change in words, not masks.
📸 Location Scout: same frame → change time of day
🎬 VFX Addition: add fire, cattle, dust — from a prompt
🧹 Set Cleanup: remove power lines, poles, anachronisms
🌪️ Atmospheric FX: add sandstorm, rain, fog to a clean plate
ALL EXAMPLES AHEAD USE YOUR FILM STILLS AS SOURCE ↓
Location Scout: Day → Night
Morero — same composition, same actor, different time of day
ORIGINAL — Golden Hour
EDITED — Moonlit Night
"Transform this exact scene to deep night. Dark indigo sky with bright stars and a low crescent moon. Grass lit by cool blue moonlight. The running man, clothing, pose, path, trees, village huts must remain exactly the same."
VFX Addition: The Angel of Life
Not Yet, Death — adding a supernatural character that doesn't exist in the source
ORIGINAL — Death at the Gate
+ ANGEL OF LIFE in Doorway
"Add a luminous Angel of Life in the doorway — radiant woman in glowing white robes, arm raised halting Death. Golden light rays across the ground. Keep Death, gate, house, sky identical."
Set Cleanup: Wire Removal
Moxaxa — remove the telephone pole and all power lines
ORIGINAL — Power Lines Visible
CLEANED — Lines Removed
"Remove the telephone pole and all power lines and electrical wires. The street lamp should also be removed. Replace with clear dark sky. Everything else — man, hoodie, shacks, muddy ground, fire, people — must remain exactly the same."
Atmospheric FX: Sandstorm
The Last Water Day — apocalyptic dust wall added behind the queue
ORIGINAL — Hazy Desert
+ DUST STORM VFX
"Add a massive approaching dust storm to the background. A towering wall of orange-brown sand rolling in from the horizon, darkening the sky. Fine dust particles blowing across the foreground. All people, woman, child, water bottles, guards, fence, sign must remain exactly as they are."
Start & End Frame → Video
Nano Banana 2 generates the end frame. Kling O3 Pro animates between them.
START FRAME — Original Still
END FRAME — Nano Banana 2 Edit
End frame prompt: "Move the man further down the path toward the village. Camera has advanced forward — more foreground path with dusty footprints. Man is smaller, mid-stride, kicking up red dust."
Start → End Frame Pipeline
Morero: The Run
Veo 3.1 animates 8 seconds between the original still and the AI-edited end frame.
AI read your treatments and imagined the key frame for every film.
PRESS ↓ TO VIEW
Tier 1 · Micro Drama
Not Yet, Death
When Death arrives early to collect a man not yet ready to die, the Angel of Life intervenes.
Written by Abale · Dir: Katamelo
▶ Video: Kling O3 Pro
Tier 2 · Flagship
The Last Water Day
In a near-future Botswana where water is rationed by law, a mother confronts an impossible choice.
Written by Nthebe Peloewetse · Dir: Pesalema
▶ Video: Kling 2.5 Turbo
Tier 1 · Micro Drama
Motho le Motho Sengwe
1985 Botswana: A university student brings her boyfriend home, sparking a hilarious family rivalry.
Written by Ruby Ramabu · Dir: Mégrangé
▶ Video: Veo 3.1
Tier 1 · Micro Drama
Skeletons in Our Closets
A reunion between childhood friends unearths secrets that challenge trust and survival.
Written by Setso Siane · Dir: Lebetha Kutlo
▶ Video: Veo 3.1
Tier 2 · Short Film
Moxaxa
A novice thug turns to robbing his community to provide for his young daughter.
Written by Molefi Moses Molefi · Dir: Larona
▶ Video: Veo 3.1
Tier 2 · Short Film
Morero
A young man races against time to find a missing cow before his lobola negotiations at sunrise.
Written by Tshepiso Monnawaletsatsi · Dir: Pius Cezuka
▶ Video: Kling O3 Pro
Tier 3 · Development
Unseen Sacrifice
A determined journalist digs into ritual murders, uncovering dark secrets linked to her sister.
Written by Lorna Bonno · Dir: Tshepo
▶ Video: Veo 3.1
Tier 4 · Pre-Colonial Epic
Polelo Yaga Serai
A young prince's destiny unravels when ancestral spirits reject his bloodline during initiation.
Written by Kago Mochotlhi · Dir: Lone
▶ Video: Kling O3 Pro
Economics of AI Pre-Viz
The Pipeline
Extract key scene from treatment
Write cinematic prompt (Claude / GPT)
Generate 4K still (Nano Banana Pro)
VFX edit (Nano Banana 2) → Video (Veo 3.1)
Cost Per Film
Generation Time: ~2 minutes
API Cost: ~$0.14
Zero stock photos.
Your story. Your IP. Your visual.
14 cents to see your film before you shoot it.
Mitigating the Risks
Own the risks before a producer asks about them.
Hallucination: Models state false info as fact. They don't "know", they predict.
Bias: Training data is Western-heavy. African stories require forceful prompting.
Privacy: Public APIs use your scripts to train future models.
The Defense: Workspaces & SkillsRAG is outdated. Agents find their own context.
Point the agent to a secure local Workspace (a folder with your scripts and budgets) and equip it with Skills (rules on how to process them). It navigates and reads what it needs.
Human review remains non-negotiable.
Your Toolkit — Free Today
Every tool below is free or has a free tier. No excuses.
📓 NotebookLM
notebooklm.google.com
Upload your syllabus, scripts, or research PDFs. Ask questions grounded in your documents — with inline citations. Generates study guides, briefings, and audio overviews. Free with any Google account.
Best for: Professors · Students · Researchers
🎬 Google Flow
labs.google/flow
Google's AI creative studio. Generate and edit images & video, control camera angles, insert/remove objects, extend scenes. Built for filmmakers. Free access via Labs.
Best for: Filmmakers · Directors · VFX Artists
🎓 Anthropic AI Courses
anthropic.skilljar.com
Free courses with certification: AI Fluency (foundations & ethics), Claude 101 (everyday tasks), Building with the API (developer track). Put the certificate on your CV.
Best for: Everyone · Free certification
⚙️ ComfyUI
comfyui.com · Open Source
Node-based visual pipeline for image & video generation. Full control over every step. Can be driven by Pi or Claude Code for automated batch workflows. The pro-level tool.
Best for: Advanced · Pipeline Builders · VFX Teams
Before You Leave This Room
One action per group. Open the URL on your phone right now.
🎬 Filmmakers
Open Google Flow → upload one still from your shoot → test insert object, extend scene, and camera control. Free.
🎓 Students
Enroll at anthropic.skilljar.com → complete AI Fluency → earn your certificate this week. Then try the script-to-call-sheet agent workflow.
📓 Professors
Open NotebookLM → upload one course syllabus + reading list → test the AI Q&A with your students' likely questions.
🏛️ Ministry Officials
Pilot an open-weights Llama 4 local deployment to protect sensitive policy data. No data leaves Botswana.
AI Studio · Fractional CTO
