The Art of Crafting the Perfect Prompt

The five strikes of a great prompt

1. Role — tell the model who it is.
2. Task — name the single outcome you want.
3. Context — give it the world it's working inside.
4. Constraints — length, tone, format, what to avoid.
5. Output shape — describe exactly what "done" looks like.

Skip a strike, lose the fight. Most weak prompts skip 3, 4, and 5.

The template

You are {ROLE with one specific specialty}.

Your task: {one sentence, one outcome}.

Context:
- {audience}
- {what already exists / inputs}
- {what success looks like}

Constraints:
- Tone: {voice}
- Length: {hard cap}
- Avoid: {hype words, hedging, lists if you want prose}

Output format:
{describe the exact structure, headings, fields, or schema}

Three sharper moves

1. Show, don't just tell

Give one short example of a "good" output. One beats ten adjectives.

2. Name the reader

"Write for a non-technical founder evaluating tools on a Sunday" beats "write clearly."

3. Force a self-check

End with: "Before answering, list what's missing from the brief. Then answer."

Anti-patterns to retire

• "Act as a world-class expert…" — empty calories.
• "Be creative." — the model will be weird, not useful.
• Stacking 12 instructions in one paragraph — split into sections.
• Asking for "the best" of anything without your criteria.

A worked example

Weak: "Write me a LinkedIn post about AI tools."

Sharp:

You are a B2B content strategist who writes like Harry Dry.

Task: Draft one LinkedIn post (max 110 words) recommending one AI tool I used this week.

Context:
- Tool: Cursor
- Audience: indie founders shipping side projects
- I want comments, not likes

Constraints:
- No hashtags
- No "in today's fast-paced world"
- One specific anecdote, one takeaway

Output format:
- Hook (1 line)
- Story (3-4 lines)
- Takeaway (1 line, contrarian if possible)

The one rule

If the model gave you a bad answer, the prompt was bad. Edit the prompt, not the output. Iterate the brief like you'd iterate a design.

— Sensei Ghost 忍