Extraction Safety Checklist
Determine whether content in a SKILL.md is safe to extract to a reference file. Used by skill authors during creation and maintenance.
SAFE to Extract (lookup/template content)
- Code templates and boilerplate blocks (>20 lines)
- Lookup tables (CWE mappings, ISTQB mappings, label definitions)
- Scoring rubrics and grading matrices
- Pattern libraries and example catalogs
- i18n tables and microcopy dictionaries
- Long checklists (>10 items) that are post-execution validation
- Format specifications (SARIF templates, OpenAPI examples)
- Consolidation summary display templates
- Directory layout ASCII trees (>10 lines)
- Multi-stack pattern comparison tables
NOT SAFE to Extract (procedural/decision content)
- Decision tables that feed directly into the next procedure step (e.g., tech stack detection in
backend-service-implement, tag routing inqa-test-generate) - Entity extraction algorithms where the table IS the algorithm (e.g.,
backend-task-breakdownentity extraction) - Workflow state machines where steps reference each other inline
- Pre-execution protocol content (already in shared reference)
- Cross-skill contract formats that downstream skills depend on (e.g., detect-dev SARIF format used by
sec-audit-remediate) - Compact tech detection tables (<10 rows, needed for every invocation)
Rule of Thumb
If removing the content would require the AI to make an extra tool call mid-procedure to look up a value it needs for the NEXT step, do NOT extract it.
How to Extract
-
Create
docs/extending/{skill-name}-reference.mdwith header:# {Skill Name} Reference Material
> Extracted reference tables and templates for the `{skill-name}` skill.
> This file is loaded by `{skill-name}` SKILL.md via inline pointers. -
Move safe content to the reference file under clear
##section headings -
Replace extracted content in SKILL.md with an inline pointer:
> **Reference**: See `${CLAUDE_PLUGIN_ROOT}/docs/extending/{skill-name}-reference.md`
> section "{Section Name}" for {brief description}. -
Verify SKILL.md is under 500 lines (target) / 600 lines (hard cap)
-
Run
scripts/validate-skills.shto confirm all gates pass
Related
- Token Strategy — Layer 2 (invocation-level) optimization
- create-skill.md — Reference Extraction Pattern section