Role: Architect & Full-Stack Engineer (solo build)

Founders usually show up with a vibe, not a spec: "an app like X, but for Y." I built Promptly to run that discovery conversation automatically. It interviews the founder, then turns the conversation into things an engineer can actually build from: modules, a dependency graph, a living PRD, and a consistency score.

The chat is the easy part. The real work is everything that stops the AI from touching the database directly.

AI brainstorming produces prose, not specs. Nothing traceable, nothing versioned.

A model that can edit your database directly means hallucinated IDs and duplicated records.

No spend cap means a surprise bill waiting to happen.

One failed AI call can leave a whole pipeline stuck.

Promptly fixes all four with actual code, not a disclaimer.

Promptly runs 16 AI agents (the Architect, Bank Planner, Extractor, Dependency Analyzer, and more), but none of them touch the database directly. Each one proposes a change through a typed tool call. Backend code checks it, and only then does it happen.

The model can never assign an ID or skip a phase. Invalid tool calls get rejected and logged, never silently applied to the database.

A founder chats with Ember, Promptly's interviewer persona, through a guided conversation. Each module gets its own question bank, 6 to 12 questions, answered by typing, picking an option, or letting the AI decide (with a confirmation step once it's decided more than half).

Every project moves through six phases: overview, modules, detailing, finalizing, review, complete. Every move is logged with who triggered it and why.

Once a module is answered, Promptly extracts real specs (title, user story, acceptance criteria) and gives each one a permanent code, like FEAT-014. It maps dependencies between them automatically. Every dependency insert gets checked against the existing graph before it's committed, so the graph can never contain a cycle. The UI and the underlying data can never disagree. The graph and the PRD document are the same data, shown two ways. Edit one, and the other updates instantly. A live consistency score tracks how solid the spec is overall.

Promptly snapshots automatically before anything risky happens: finalizing, reopening, restoring. You can roll back a single module or the whole project. Even restoring has its own safety net, since every restore takes a snapshot before it runs.

Four WebSocket rooms per project keep chat, graph, document, and score in sync live, backed by Redis so it works across multiple servers, not just one. Sessions use rotating refresh tokens with reuse detection. If an already-used token gets presented again, the entire session chain gets invalidated, cutting off whoever stole it. Every one of the 16 AI pipelines has a fallback built in. If a step fails, the user never gets stuck staring at a dead spinner.

Backend: Python, FastAPI, SQLAlchemy, PostgreSQL Real-time: Redis, WebSockets AI: Anthropic API (Claude), streaming + tool calls Frontend: Next.js, React, TypeScript, Tailwind

The AI interviews you. The server owns every decision it proposes.

That's the whole idea behind Promptly. The model gets to be creative. It never gets to be in charge.

I built it because I kept seeing the same breakdown happen over and over. A founder says one thing, the developer hears something else, and three weeks in you find out the whole build was based on a misunderstanding nobody caught early. Promptly exists to close that gap. The interview forces the vague idea into specific, written answers before a single line of code gets planned, so what the founder meant and what the engineer builds actually match.