cowork plugin skill —
skills/only the category is used (no commands/·agents/).
6 core principles (Claude official)
- Be clear and direct — Treat Claude as a 'smart new colleague with no context.' Directly specify the desired output format and constraints. "Write me some help" (X) → "5-step installation help for beginners, 2 sentences per step, 1 emoji" (O).
- Add context and motivation — 왜 If you explain why it matters, Claude generalizes correctly. "I'll be reading it aloud, so no ellipses" > "No ellipses."
- Show with examples (multishot) — 3–5 well-crafted examples most reliably guide format, tone, and structure.
<example>Wrap in tags. Positive examples are stronger than negative instructions. - Structure with XML tags — For mixed prompts,
<instructions>,<context>,<input>,<output_format>splitting them with tags keeps Claude from getting confused. - Assign a role — A one-sentence role in the system prompt ("You are a copywriter with 10 years of experience") focuses tone and behavior.
- Put long materials at the top — For 20k+ token inputs, place the data at the top, and the questions and instructions at the bottom(quality improves ~30%).
Practical cheat sheet for non-developers
| Requests like this | Change them to this |
|---|---|
| "Write me an article" | "A health insurance blog post for housewives in their 50s, 3 paragraphs, 1 key benefit" |
| "Make it good" (vague) | Show 1 example of the desired format yourself |
| "Don't do ~" (negative) | "Write it like ~" (positive) — negative instructions are poorly followed |
| "Make it better" (evaluative word) | Specific criteria ("understandable by a 5th grader, 25 characters or fewer per sentence") |
Frequently asked questions
Q: Claude writes too long → Specify the output format: "3 paragraphs, 25 characters or fewer per sentence" or "1 key line per slide."
Q: My nuance doesn't come through → Provide a role + 2–3 examples. "In this tone: [example 1] [example 2]."
Q: It lies (hallucinates) → "If you don't know, say you don't know. Don't guess" + request verifiable sources.
Reference
- Claude official prompt engineering guide: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
- This skill is
/harness:plugina cowork plugin skill generated by the harness (skill-builder) based on Claude's official documentation.