progressive-disclosure

To reduce context-window issues, progressively disclose large agent skills by splitting verbose instruction sections into smaller, directly linked reference files while preserving the original guidance and meaning.

Progressive disclosure for YAML frontmatter and ACP definitions is a refactoring technique that moves detailed skill instructions into linked reference documents, allowing agents to load context dynamically without exceeding token limits.

Split DataRobot skill instructions by applying safety-first editing guardrails that move content minimally into reference files, strictly preventing rewrites that alter intent, remove critical prerequisites, or change SDK command guidance.

You need to chunk skill instructions into separate reference files when a skill triggers context warnings, suffers from context-window limits, or benefits from progressive disclosure during agent execution.

Limitations include the strict requirement to keep links one level deep from SKILL.md to references, and the absolute necessity of running validation and lint tasks to ensure no semantic changes occurred during the documentation refactor.

Progressive disclosure works with any agent skill that triggers context warnings, needs chunking, or benefits from progressive disclosure during agent execution, although it is specifically designed to refactor DataRobot skill instructions.