A domain composition practice

AI Origami

AI Origami is the practice of folding knowledge artifacts, tools, workflows, skills, connectors, and operating context into a domain-capable AI workbench.

The goal is not a better chatbot. The goal is an agent that can work inside a bounded technical domain with the references, conventions, and live context the job requires.

See the practice View the MTP example
The model is the sheet. The structure around it determines what it can become.
The workbench is the specialist. Domain competence is composed into the working environment.

Why it exists

Generic fluency is not domain competence.

General AI can explain a technical domain. That is useful, but it is not the same as being able to operate inside one.

In industrial automation, useful work depends on standards, vendor behavior, project conventions, tested examples, current system state, and the judgment that accumulates across real jobs. AI Origami starts from the premise that those materials should shape the agent's working environment, not sit outside the conversation as scattered references.

  • Not trained into the model The domain is engineered into the place where the model works.
  • Not just documents in context The workbench includes tools, permissions, workflows, examples, and operating rules.
  • Not just static knowledge When the domain depends on current state, connectors bring live data into the workbench.

The precursor ladder

Each fold makes the workbench more capable.

AI Origami sits at the top of a practical stack. Each layer gives the agent more structure for doing real work in a specific domain.

Base Model Language understanding, reasoning, and generation.
Harness Engineering A structured place for the agent to run.
Connectors Access to tools, data, and external systems.
Skills Reusable know-how packaged for repeated use.
Workflows Domain job sequences that chain skills into work.
AI Origami The composed domain-capable workbench.

Canonical example

The MTP workbench made the distinction visible.

Module Type Package work is not hard because one document is hard to summarize. It is hard because useful work lives between standards, supplier documentation, PLC implementation details, SCADA tooling, project conventions, and live system behavior.

In the MTP example, a general-purpose agent became useful because the environment was composed around the job. The workbench could navigate vocabulary, generate reference material, support design decisions, and assist with troubleshooting against live SCADA data.

  • Standards NAMUR and MTP reference material This gives the workbench the formal vocabulary and rules of the domain, so it can distinguish a plausible MTP answer from one that fits the standard.
  • Supplier context Festo PEA documentation The supplier material turns abstract MTP ideas into the behavior of a real process equipment assembly, including quirks the standard alone will not explain.
  • Implementation Codesys project details PLC structure anchors the conversation in actual control logic, so the agent can connect MTP concepts to signals, blocks, and project files engineers must maintain.
  • SCADA context WinCC OA material and examples SCADA references show how the package is imported, displayed, and operated, which is where many integration problems become visible.
  • Connector Live SCADA access through MCP Live access changes the workbench from document reasoning to system reasoning. It can compare an explanation against current plant-floor state.
  • Project memory Conventions, examples, and libraries Local examples tell the agent what good work looks like here, not just what generic documentation says should be possible.

Boundaries

What AI Origami is not.

Not fine-tuning

The model is not retrained to become an internal expert.

Not RAG alone

Retrieval may be one fold, but composition is the practice.

Not a chatbot wrapper

The workbench includes tools, workflows, permissions, and context.

Not a product

This is a developing practice for domain-capable AI work.

Project status

A working concept, still being shaped.

AI Origami is currently a concept-development project rooted in industrial automation practice. This site is the public surface for the framework: enough context to understand the practice without exposing the working notes behind it.

  • Current public claim Domain capability can be composed into the agent's working environment without retraining the model.
  • Evidence so far The MTP workbench shows how standards, supplier context, project memory, and live system access turn generic fluency into bounded domain work.
  • Next public expansion Define the folds more precisely, document additional examples, and clarify how this differs from RAG, fine-tuning, and chatbot wrappers.