Introducing INKBOT: Building an Intermediate Intelligence Architecture for Human Intent

Description and use

INKBOT introduces an intermediate intelligence layer that lets users explicitly define and correct intent before model execution, reducing friction between human meaning and model inference.

Problem

Multimodal AI systems struggle to align model inferences with subtle human intent, creating friction when users must act as prompt engineers.

Outcome

INKBOT provides a local-first web file architecture that packages human-approved intent separately from model outputs, enabling multi-step workflows and reducing alignment friction.

Workflow

• Describe the desired outcome
• Make the intent visible in a structured format
• Recognize model inferences against the intent
• Correct mismatches
• Refine the intent and repeat if necessary

Evidence

I kept hitting a wall with multimodal AI systems: the friction between what a human naturally means and what a model infers.

Source

Instead of forcing the user to become a prompt engineer, I wanted to see if we could build an intermediate intelligence architecture layer to make human intent reviewable and corrigible before the model executes a final build.

Source

The loop I’m playing with is: Describe → Make it Visible → Recognize → Correct → Refine.

Source

The architecture sits entirely in a single local-first web file.

Source