PWTL (Prompt Workflow & Task Language)

The problem

AI is asked to write answers. When it doesn’t know, it writes a convincing one anyway, and nothing in the output tells you which answers were guesses. Prompt injection makes it worse: a single uploaded file can quietly change what the model does. That’s why most AI never gets trusted with decisions someone has to sign.

The solution

PWTL is a task language for AI judgment. Instead of asking a model to write an answer, it asks the model to choose one, and to show its work. I built it for GARDE1, where every decision has to hold up in front of an assessor.

Without PWTL	With PWTL
The model writes whatever sounds right	The model chooses among outcomes defined before it was asked, so it can’t invent one
You can’t tell what an answer was based on	Every decision cites the evidence it relied on
The reasoning disappears	The reasoning is kept, including what was ruled out and why
An uploaded file can hijack the model	Uploaded content can inform a decision but never instruct the model
Bad answers ship	Answers are checked before they count
Long prompts bury the facts that matter	The model sees the facts that matter and little else

What changes

AI judgment you can audit, which is the only kind a regulated business can use. PWTL runs everywhere GARDE1 asks a model anything: evaluating assessments, writing and checking documents, and deciding whether a piece of evidence proves anything at all. That’s why the platform can refuse an upload that proves nothing, and say which claim it was for.

It works anywhere a decision has to be defensible: compliance, insurance, healthcare, finance, code review.

Technical deep dive available under NDA. Get in touch.