phase-workflow-skill-creator

To decompose a monolithic agent prompt into a phased workflow, split the content into a lightweight router SKILL.md and per-phase reference files. This progressively discloses context and enforces transition gates between execution stages.

Large agent prompts become unwieldy to maintain because monolithic files load excessive context simultaneously. Splitting them into a router and granular phase reference files reduces context load and improves reasoning.

The best way to enforce transition gates during staged agent execution is to structure the workflow with completion checks before moving between the Read Workspace, Core Work Phases, and Produce Output stages.

Structure reference files for progressive context disclosure by keeping the agent file slim with only core roles, I/O contracts, and universal rules, while moving granular domain content into separate phase references loaded only when needed.

Yes, this phased workflow approach works for long running prompts by converting bloated agents into phase-driven workflows with reusable references, ensuring staged execution and progressive context loading.

You should not use a router-based phase navigation approach for simple agents that do not require progressive context disclosure or staged execution, as the overhead of managing transition gates and reference files outweighs the benefits.