ScenarioForge AI

Inspiration

Most professional training programs share the same fundamental flaw: you read a course, watch a video, pass a multiple-choice quiz — and then face a real client, a difficult manager, or a high-stakes negotiation with no real sense of how you'll perform. The gap between knowing and doing is where careers stall.

Traditional role-play exercises exist, but they require a human trainer, scheduled sessions, and subjective feedback. For the millions of learners who can't access executive coaching or structured practice environments, there is simply nothing. ScenarioForge AI was built to close that gap — making high-quality skill practice available to anyone, at any time, on any topic.

What it does

ScenarioForge AI is an interactive learning simulation platform where anyone can practice real-world professional skills through immersive AI-powered conversations.

A learner picks a scenario — salary negotiation, medical diagnosis, legal cross-examination, closing a resistant client — and is placed in a live dialogue with an AI character that responds in-role, adapts dynamically to what they say, and challenges them the way a real person would. No scripted paths. No multiple choice. A real conversation with real stakes.

At the end of every session, the platform generates a detailed animated debrief: a circular score gauge filling on screen, a star rating, and section-by-section color-coded feedback — green for what went well, red for areas of improvement, blue for three concrete tips — all referencing specific moments from the actual conversation. If the AI escalated its resistance mid-session because the learner was performing too well, and the learner handled it, the debrief explicitly recognizes it.

Trainers and educators can also build their own scenarios in minutes using the Creator Mode wizard, or start instantly from a library of 12 pre-built templates across 4 professional domains.

How we built it

ScenarioForge AI was built entirely through iterative conversations with MeDo, without writing a single line of code manually.

The process followed a layered approach. We started by describing the core concept — a split-screen chat interface with a live feedback panel — and MeDo generated the full React/TypeScript front-end with a Supabase backend, including the database schema for scenarios and sessions.

From there, each feature was added through focused multi-turn prompts: the Creator Mode wizard, the AI character opening message system, the live confidence scoring engine, the adaptive difficulty escalation logic, the objectives completion detection, and finally the pre-recorded demo mode. MeDo also generated the complete serverless architecture — a Supabase edge function that receives the full conversation history, character configuration, and success criteria, then returns both the AI's in-role response and a structured metrics update in a single call.

Challenges we ran into

The hardest challenge was the real-time scoring system. Getting the AI to simultaneously act in-role and evaluate the learner's performance at each exchange — returning structured metrics alongside a natural dialogue response — required significant prompt engineering and iterative refinement through MeDo's multi-turn chat.

The adaptive difficulty escalation was equally complex: defining the exact threshold logic, capping escalation at 2 events per session, and ensuring the AI character's tone shifted convincingly without breaking immersion required several rounds of testing and adjustment.

The demo mode also presented an unexpected challenge. MeDo's first interpretation was to reuse the existing pre-built scenarios as a live demo. We had to be very precise in re-describing what a pre-recorded session playback actually means — static hardcoded data, zero API calls, purely animated — before the correct implementation was produced.

Accomplishments that we're proud of

We are most proud of the debrief experience. What could have been a simple score summary became a genuine coaching moment — animated, personalized, and grounded in the actual conversation. Every bullet point reflects something that truly happened in the session, making feedback feel earned rather than generated.

We are also proud of the adaptive difficulty system. The idea that an AI character responds to a learner's performance in real time — becoming harder when the learner is excelling, and acknowledging that escalation explicitly in the final debrief — creates a training loop that feels alive and genuinely challenging.

Finally, building a full-stack production application with database, serverless functions, real-time UI, and 12 content-complete scenario templates entirely through natural language conversations with MeDo is itself an accomplishment worth noting.

What we learned

We learned that precision in prompting is everything. Vague descriptions produce working but generic results. The features that impressed us most came from prompts that described not just what to build, but why it should behave a certain way — the user experience intent behind the technical requirement.

We also learned that MeDo's multi-turn architecture is best leveraged as an iterative design tool: build a foundation, test it, identify what's missing, and refine layer by layer — rather than trying to describe the entire application in one prompt.

What's next for ScenarioForge AI

The immediate next step is voice interaction — replacing the text input with speech recognition so learners can practice negotiation or medical interviews out loud, the way these conversations actually happen in real life.

Beyond that, a Trainer Analytics Dashboard would allow educators to track learner performance across multiple sessions — seeing which objectives are consistently missed, which scenarios are most challenging, and where cohorts need focused coaching.

Longer term, ScenarioForge AI could integrate with existing Learning Management Systems and serve as the practice layer that every course is currently missing: the space between theory and reality where real skill is built.

Built With

• medo
• react
• supabase
• typescript