How it works
Describe. Decide. Build.
Three steps. One conversation. No guessing.
The Dialog
You describe what you want. Pantion doesn’t accept it at face value.
It asks one question at a time. Each question targets a specific ambiguity — not filler, not process, just decisions that need to be made. You answer in your own words.
Every answer gets classified:
HARD — invariant. “Users must confirm before deleting data.” This propagates to every agent, every rebuild, forever.
FLEX — safe default. “Pagination defaults to 25 items.” This can evolve without breaking anything.
If something is unclear, it becomes an open question — tracked, numbered, resolved before the dialog ends. Nothing gets silently guessed.
The dialog continues until convergence: the point where new questions stop producing new behavior. When there’s nothing meaningful left to ask, you’re done.
The Canon
The converged dialog becomes your canon — the source of truth for what you actually meant.
It preserves what a summary loses: rejected alternatives, the order in which decisions were made, the reasoning behind each choice. A future model can re-read it and understand not just what you decided, but why.
After you approve the canon, Pantion translates it into spec files your agent can build from. Every spec traces back to a specific decision in the dialog. Nothing appears without a source.
The canon is not implementation
No framework. No database. No tech stack. The canon describes behavior — what the system does, not how it’s built.
This means you can rebuild from the same canon in a completely different stack. The intent doesn’t change.
The Build
Your agent reads the canon and builds what it describes.
The difference from a normal prompt: the agent doesn’t have to guess about edge cases, defaults, permissions, or failure behavior. Those decisions were already made — by you, in the dialog.
And when requirements change? You amend the canon. The original stays intact. The change is explicit, traceable, and rebuildable.