status: active
timestamp: 2026-07-02
tags: [agent, skills, composition, modularity]
Small composable skills, not mega-skills
One skill does one thing well. Chain 4-6 small skills > one 500-line mega-skill. Enables auto-invocation per sub-task + composition across workflows.
Small composable skills
The rule
Each skill does ONE thing. Not “write my whole YouTube video” — instead: script-critique, intro-writer, title-generator, description-writer. Four skills, each focused, all invokable independently.
Why
Auto-invocation: Claude auto-loads the right skill based on the user’s phrasing. write me a title → title-generator. critique this script → script-critique. A mega-skill sits idle until explicitly named; small skills fire when they match.
Reusability: title-generator works for videos AND blog posts AND emails. A full-youtube-workflow skill only works for YouTube.
Debuggability: When a skill produces wrong output, you know exactly which of the 6 skills to fix. In a mega-skill, the failure mode is opaque.
Composition: Real workflows chain skills. Nate Herk 2026-07-02: script → critique → intro → title → description = 4 skills in one project. Not one skill doing all 4.
The shape
| Good skill | Bad mega-skill |
|---|---|
script-critique.md — 60 lines | youtube-full-workflow.md — 400 lines |
intro-writer.md — 40 lines | |
title-generator.md — 30 lines | |
description-writer.md — 30 lines |
Total: 160 lines across 4 skills > 400 lines in one.
Size heuristic
If a SKILL.md gets past ~150 lines, ask: can this split into 2 skills? If yes, split.
Exceptions:
- Reference/lookup skills (e.g.
github-actions-docs) — inherently long because they encode documentation - Deep-domain skills (e.g.
webapp-testing) — the domain is genuinely broad; split by sub-domain if possible but don’t force it
What “one thing” means
- ✅
code-review— reviews code for defects - ✅
security-review— reviews code for OWASP-style vulns - ❌
code-and-security-review— do both
Related-but-separate concerns get separate skills. The user can invoke both in sequence.
When to combine
When to actually merge two skills into one:
- Both skills share ≥80% of their prompt content
- Neither skill is useful alone (one always fires with the other)
- Combined skill is still <150 lines
Otherwise, keep separate.
Cross-refs
ponytail§ULTRA — “one implementation = no interface. one product = no factory”iterate-before-skill— iteration reveals the natural boundaries between skillsdelegate-to-subagents-by-default— a mega-skill vs many small skills is the same anti-pattern as monolithic agent vs subagent orchestration- Source: Nate Herk YouTube 2026-07-02 — his 4-skill YouTube workflow demonstrates composition