# Model context and tool-call efficiency

The MCP initialization response carries concise shared operating rules. Detailed workflow examples remain available through the existing `dominus-workflow` prompt. Dedicated tool descriptions still carry their own inputs, constraints, and failure semantics; the startup instructions do not require retrieving the full handbook for every task.

The September 7 comparison measured 4,647 characters in startup guidance versus 10,774 in the previous handbook, a 57% reduction in that text. This is a character measurement, not a tokenizer benchmark or a claim about total session cost. Tool schemas, messages, images, model choice, and host behavior all affect actual token usage. The handbook can grow independently of startup guidance.

MCP integration checks verify the negotiated startup guidance, continued prompt retrieval, and a size budget relative to the handbook. They retain revision locking, managed QA guidance, and key-handling rules in startup context. These checks do not establish that every MCP host displays or follows server instructions.

Other implemented reductions include local-source omission when synchronized, incremental output cursors, batch commerce creation, batched scene checks, internal runtime discovery, and check/input/check QA scenarios. These reduce duplicated data or model round trips; internal Studio and network operations still occur. Live workflow measurements and regression scenarios remain release work.