Meta-Skill-Engineering skill-variant-splitting

Purpose

Split a skill that has grown too broad into focused variants that route precisely and perform well. A skill covering too many cases develops vague triggers, bloated procedures, and diluted output.

When to use

• User says "this skill does too much", "split this skill", "create variants"
• Skill has disjoint sections ("For Python:" / "For JavaScript:")
• Skill triggers on unrelated inputs causing routing confusion
• Procedure has conditional branches per input type
• Description exceeds two sentences trying to cover all cases

Do NOT use when:

• Skill is focused and working well
• Variations are minor enough for overlays
• Problem is trigger wording only →

  skill-trigger-optimization

• Porting a working skill to a different context →

  skill-adaptation

• Reorganizing skills at catalog level →

  skill-catalog-curation

• Skill genuinely handles one coherent task

Procedure

1. Identify splitting signals

   • Count distinct "modes" — if/else branches, "For X:" sections
   • Count non-overlapping trigger clusters
   • Check whether output format varies by case
   • Check whether different contexts need different tool access or behaviors

2. Determine split axis

   • Stack: e.g.

     skill-testing-python

     ,

     skill-testing-javascript

   • Platform: e.g.

     skill-deploy-aws

     ,

     skill-deploy-gcp

   • Scope: e.g.

     skill-review-quick

     ,

     skill-review-thorough

   • Domain: e.g.

     skill-api-rest

     ,

     skill-api-graphql

   Axis selection priority (when multiple axes seem viable):

   1. Prefer the axis that produces the fewest variants while still eliminating conditional branches.
   2. If two axes produce the same variant count, prefer the axis whose variants have the most distinct trigger vocabularies (easier routing — fewer false positives).
   3. If still tied, prefer: Stack > Domain > Platform > Scope. Stack splits tend to have the cleanest boundaries because different stacks use different terminology.
   4. Red flag: If any axis produces >5 variants, the skill may need hierarchical organization (umbrella router + sub-skills) instead of a flat split.

3. Define variants

   • Each variant passes the "one sentence" test — scope fits in a single clear sentence
   • Each has distinct, non-overlapping triggers
   • Each has a focused, linear procedure without conditional branches

4. Extract shared core (if any)

   • Identify logic common across all variants
   • If shared content exceeds 30% of each variant's total length, create a base skill with extensions
   • If shared content is under 30%, create fully independent variants
   • If creating a base skill: put shared procedure in the base, variant-specific steps in each variant's SKILL.md

5. Write each variant

   • Name:

     {original}-{variant}

   • Action-verb-first description
   • Simplified procedure
   • Variant-specific examples

6. Update routing

   • Each variant's "Do NOT use when" references siblings by name
   • Example: "Do NOT use for GraphQL APIs →

     api-testing-graphql

     "

7. Verify coverage

   • Every original use case handled by exactly one variant
   • No gaps between variants
   • No overlapping triggers

8. Deprecate or repurpose original

   • Fully replaced → deprecate
   • Some general cases remain → keep as lightweight router

Output contract

Produce a markdown document with this structure:

## Variant Split: [original skill name]

### Split Axis
[Stack | Platform | Scope | Domain]: [one-line reasoning]

### Variants
| Variant | Scope | Key Triggers |
|---------|-------|-------------|
| original-a | [focused scope] | "trigger 1", "trigger 2" |
| original-b | [focused scope] | "trigger 3", "trigger 4" |

### Shared Core
[Shared logic description, or "None — fully independent variants"]

### Coverage Map
- [x] Case 1 → variant-a
- [x] Case 2 → variant-b
- [x] No gaps or overlaps confirmed

### Migration
Original skill: [deprecate | keep as router | keep for general cases]

Failure handling

• No clean axis found: Skill may be unified after all. Report that no beneficial split axis exists and recommend

  skill-improver

  instead.
• Variants overlap on triggers: Rethink the axis. If overlap persists, document the ambiguous boundary and escalate the decision to the user.
• More than 5 variants: Use a two-level hierarchy with an umbrella router skill instead of a flat split.
• Shared core larger than variant-specific content: Splitting adds duplication without benefit. Recommend refining the original instead.

Next steps

After splitting:

• Update the library index →

  skill-catalog-curation

• Evaluate each variant →

  skill-evaluation

• Deprecate the original if fully replaced →

  skill-lifecycle-management