When to Use Skills vs. Simple Prompts

When to Use Skills vs. Simple Prompts | Generated by AI

Home 2026.05

Question: When should you actually create a “skill” vs. just saving a prompt or using Claude Code directly? Is skill creation only for hard/complex tasks?

Answer:

Yes, you’ve got the right mental model. Here’s the clear breakdown:

When to create a Skill

A skill is justified when the task is complex enough that a naive LLM call fails, and you need:

Multi-step orchestration — the task requires a sequence of tool calls, file reads, bash commands, and decisions chained together
Environment-specific constraints — e.g., “use --break-system-packages”, “write output to /mnt/user-data/outputs/”, “this library has a quirky API” — things that aren’t in the model’s training data
Repeatability at quality — you’ve already trial-and-errored the right approach and want to encode that knowledge so every future run gets it right the first time
300+ line implementations — when the “how to do it correctly” is itself non-trivial code (your ww tool is a good example of this threshold)

Skills are basically compiled institutional knowledge — the delta between “naive Claude attempt” and “Claude that already did this correctly 10 times.”

When NOT to create a Skill — just save a prompt

For tasks where a well-written prompt in a markdown file is enough:

~/prompts/
  summarize-pr.md
  review-python.md
  explain-paper.md
  draft-commit-msg.md

Then cat ~/prompts/summarize-pr.md | pbcopy or pipe it directly into your CLI agent. This covers:

Consistent formatting requests — “always respond in this structure”
Domain-specific context injection — “assume the codebase is Django + Celery + Postgres”
Persona/style constraints — “be terse, code-first, no preamble”
Simple one-shot tasks — anything Claude gets right on first try with the right context

The actual heuristic

Does naive Claude (no extra context) fail this task?
├── No  → Just prompt. Save it in a .md file if you repeat it.
└── Yes → Does it fail because of missing *knowledge* or missing *orchestration*?
    ├── Missing knowledge (env quirks, library gotchas, your conventions)
    │   → Add a SKILL.md with the encoded know-how
    └── Missing orchestration (multi-step, tool chaining, stateful)
        → Build a script (like ww) OR a SKILL.md that guides the orchestration

Your `ww` tool is already the right instinct

ww at 200-300 lines is exactly the right level — it’s a reusable agent harness where the complexity lives in the Python, not in a skill file. A SKILL.md for that would just be:

# ww-style CLI agent
Run: python ww.py <task>
The script handles tool dispatch, context injection, and retry logic.
Don't reinvent — extend ww.py directly.

The skill file just points to the real implementation.

TL;DR

Situation	Solution
Repeating a prompt pattern	Save `.md` prompt file
Claude fails due to env/library quirks	SKILL.md with constraints
Complex multi-step orchestration	Python script (like `ww`)
Both quirks + orchestration	SKILL.md + script combo
Simple Q&A	Just ask, don’t save anything

Skills are worth the overhead only when the encoded knowledge meaningfully improves output quality on a task you’ll run repeatedly. For everything else, a prompt file or direct Claude Code invocation is the right call.

Back Donate