CASE STUDY · MY PRODUCT · AI SAAS
PromptMotion
Turns a creator's rough idea into production-ready prompts for AI video models — powered by a multi-agent engine I designed, built, launched in 5 months and grew to 1,500+ users in its first 2 months.
CREATOR TYPES
“a lonely robot finds a flower in a ruined city”
One line. No camera language, no lighting, no continuity plan.
PROMPTMOTION OUTPUT
generating
- MY ROLE
- Founder · engine architect
- USERS
- 1,500+ in 2 months
- MODEL
- Free tier + paid plans
- STATUS
- Live in production
THE PROBLEM
Creators know the video they want. They don't speak the language AI video models need.
Camera, lighting and motion terms most creators don't know.
Characters and places drift between clips.
Veo, Kling and Runway each want prompts written differently.
Every bad prompt costs a paid generation.
From idea to a shootable plan.
A two-stage result page: lock the characters first, then generate each shot, with a guided dock showing what to do next.
The AI plans camera, lighting, movement and composition before it writes the final prompt — the way a real director would.
The same plan is written natively for Veo 3.1, Kling, Runway, MiniMax and Seedance.
Finished videos made by creators using PromptMotion prompts.
From one LLM call to a team of agents.
Each step was driven by a real limitation users hit — not by chasing frameworks.
V1
Single LLMOne big prompt. Fast to ship, but shallow and inconsistent.
V2
Multi-agentSplit into specialist agents, each with one job.
V3
RAG knowledgeA cinematography knowledge base agents retrieve from.
V4
Agents & criticsShared continuity ledger and a critic loop that sends work back.
V5
CachingCaching infrastructure to bring the cost per run down.
Under the hood: the Script Architect
Watch it run live ↓AGENT 1
Concept AnalystFinds the emotional core
AGENT 2
Story ArchitectBuilds the arc and cast
AGENT 3
Scene WriterDesigns each shot
AGENT 4
Continuity CheckerChecks against the ledger
AGENT 5
Output FormatterWrites per-model prompts
Append-only state shared across every clip, so cast and places never drift.
Retrievable technique-to-emotion mappings, instead of stuffing them into prompts.
Drafts are reviewed and sent back with specific feedback until they pass.
Built with LangChain and LangGraph.
5–10 agents that plan, debate, check and critique each other.
A full run on a real brief — every handoff, knowledge-base lookup, question and critique, as it happens.
brief: “a lonely robot finds a flower in a ruined city”
handoff · 1/25
1
MESSAGES
0
KB LOOKUPS
0
MEMORY OPS
0
CRITIQUES
0
REVISIONS
AGENT DISCUSSION
phase: ideation
- Director→ Concept AnalystHANDOFFNew brief: “a lonely robot finds a flower in a ruined city.” Find the emotional core.
Handoff
RAG retrieval
Shared memory
Discussion
Critique
Revision
Approval & delivery
THE KEY LESSON
Consistent isn't the same as good.
The engine became reliable — characters stayed on-model, shots stayed coherent. But it was optimised for consistency, not for “wow”. So I redesigned the loop to reward ideas that surprise.
01
Vision CardA “taste DNA” artifact every agent works from.
02
Idea gateSeveral premises, scored for how divergent they are.
03
Divergence engineDirector-lens seeds push ideas in different creative directions.
04
Two criticsOne judges correctness, the other ranks for taste.
The hard parts.
01
Cost of many agents
More agents means more LLM calls. A caching layer brings the cost per run down.
02
Continuity across clips
Generated clips are independent; the ledger is what keeps them one story.
03
Moving model targets
New video models keep launching, each with its own prompt dialect.
04
Measuring “good”
Correctness is checkable. Taste needed its own critic and scoring.
Engineering is half of it.
I own the product end to end — pricing, acquisition, activation and teaching users to succeed.
Free tier plus paid subscriptions, priced in LKR for the home market.
Organic search did most of the work, plus a small Meta budget and localised ad creatives — no Google Ads.
Meta Pixel + Conversions API with deduplication, and a custom “first prompt” activation event.
Creator Academy community with recorded tutorials, including Sinhala workflows.
Results
1,500+
Users in the first 2 months
5
Video models supported
5 mo
From idea to launch
5–10
Specialist agents per run, depending on the brief
My role
- Founder — product, pricing and go-to-market
- Designed and built the multi-agent engine
- RAG knowledge base and caching layer
- Product UX, analytics and ads infrastructure