| name | progressive-disclosure |
| description | Refactor large DataRobot skill files by moving detailed content into directly linked reference files while preserving meaning. Use when a skill triggers context-window warnings, needs progressive disclosure, or should be chunked without changing guidance. |
Progressive Disclosure
Use this repo-local meta skill to split large skills into a concise SKILL.md plus adjacent reference files. This guides content moves and reference creation; it is not a content rewrite pass.
Guardrails
- Do not rewrite, reinterpret, simplify, or change the meaning of skill content.
- Do not delete details because they seem verbose. Move details to reference files instead.
- Do not change SDK guidance, commands, workflows, safety notes, or examples except to preserve links after moving content.
- Do not edit plugin manifests, versions,
CODEOWNERS, or unrelated repo metadata unless the user explicitly asks.
- Do not create deeply nested references. Reference files should be directly linked from
SKILL.md.
- Do not move first-use trigger guidance, required safety constraints, or critical prerequisites out of
SKILL.md.
- If content needs substantive editing, stop and ask the user before making that change.
Target Shape
Keep SKILL.md focused on:
- Frontmatter with accurate
name and description
- When to use the skill
- Quick start or primary workflow
- Key decision points and safety constraints
- Direct links to reference files
Move long supporting content into files such as:
reference.md
examples.md
troubleshooting.md
platform-setup.md
sdk-operations.md
ci-cd-patterns.md
Choose names that match the moved section. Prefer a small number of meaningful files over many tiny fragments.
Use folders only when the supporting material has a clear type or grouping:
references/ for longer conceptual guides, examples, and workflow details
scripts/ for executable helpers that the skill may ask the agent to run
- Language folders such as
python/, typescript/, or r/ when a skill supports multiple programming languages
Keep these folders directly under the skill directory. SKILL.md should link directly to files inside them, such as references/workflow-details.md, without requiring the agent to follow a chain of references.
Workflow
-
Inspect the target skill:
- Read
skills/<skill-name>/SKILL.md.
- Check current line count and major headings.
- Check current token warning/error status from
task test:integration when available.
- Identify sections that are detailed reference material, long examples, command recipes, troubleshooting tables, platform-specific variants, or API details.
-
Propose the split before editing when the change is substantial:
- List which sections will stay in
SKILL.md.
- List which sections will move and the destination reference file for each.
- Keep a simple traceability map: original heading -> destination file -> replacement link text.
- Confirm any ambiguous sections with the user.
-
Move content with minimal transformation:
- Preserve headings and body text as much as possible.
- Adjust heading levels only when needed for valid Markdown structure.
- Preserve code blocks exactly unless a relative link/path must change.
- Keep warnings, prerequisites, and safety notes attached to the relevant content.
-
Replace moved content with concise pointers:
For detailed deployment configuration, see [deployment-reference.md](deployment-reference.md).
Each pointer should explain when to read the reference, not summarize the whole reference.
-
Keep links one level deep:
SKILL.md should link directly to every reference file the agent may need.
- Reference files may link back to
SKILL.md if useful.
- Avoid chains like
SKILL.md -> references/overview.md -> references/deep-reference.md.
-
Validate:
- Run
task test:integration for structural checks.
- Run
task lint when tooling is available.
- If the goal is reducing context-window warnings, rerun the relevant tests and compare warning output. This repo warns at an estimated 3300 tokens and errors at 6700 tokens.
Candidate Section Heuristics
Good move candidates:
- Long examples that are not needed for initial task routing
- Repeated setup variants by platform or provider
- Detailed SDK method inventories
- Troubleshooting catalogs
- Extended CI/CD snippets
- Long validation or debugging procedures
Usually keep in SKILL.md:
- The first-pass decision tree
- Required safety constraints
- Minimal happy-path workflow
- Short command checklist needed for most runs
- Links to all supporting references
Diff Review Checklist
Before finishing, verify:
- Every moved block still exists in a reference file.
SKILL.md links directly to each new reference file.
- The frontmatter
description still includes clear trigger scenarios.
- The final response includes the traceability map for reviewer confidence.
- No instruction, recommendation, or example changed.
- No unrelated formatting churn was introduced.
- Repo validation passes or any local tooling gap is clearly reported.