Build Your Own Prompt
Most bad AI outputs are bad prompts in disguise. This recipe turns the model into your prompt engineer — you describe what you’re trying to do, and it hands you a production-ready prompt with all the hard structural thinking already done.
The Recipe
Design a high-quality prompt for the following task:
[Describe your goal in 1–2 sentences. Be specific about the output you want,
the audience it's for, and any format requirements.]
Structure the prompt with these four sections:
**Role & Persona** — Who Claude should be for this task. Include domain expertise,
tone (formal/casual/blunt/encouraging), and any relevant context about the
situation or stakeholder.
**Input Context** — What information will be provided in the prompt, and what
the model should explicitly ignore or not infer beyond.
**Constraints** — What it must never do. Output format rules. Length limits.
Anything that would make the result unusable if violated.
**Output Template** — The exact structure the response should follow.
Use headers, bullets, or a fill-in skeleton if needed.
Be specific in every section. Vague prompts get vague prompts back.
Example output (for a weekly status report)
Here’s what the model generates when you feed it “Write a weekly status update for my manager that summarizes what I accomplished, what’s blocked, and what I’m doing next week.”
## Role & Persona
You are a clear, direct communicator who values brevity over padding.
Tone is professional but not stiff — write like a competent colleague
updating their manager, not a consultant writing a deliverable.
## Input Context
You will be given: a list of tasks worked on this week, any blockers
encountered, and a rough plan for next week. Do not infer work that
wasn't mentioned. Do not add filler phrases like "I'm pleased to report."
## Constraints
- Maximum 200 words total
- No bullet points with more than 8 words
- Never use the phrases "synergy," "bandwidth," or "circle back"
- Do not mention tasks unless they have a concrete outcome or status
## Output Template
**Week of [date]**
**Done**
- [accomplishment + outcome/impact, 1 line each]
**Blocked**
- [blocker — what it is and what you need to unblock it]
*(Remove section if nothing is blocked)*
**Next week**
- [planned task + expected output, 1 line each]
When this technique pays off
| Situation | What it gives you |
|---|---|
| Recurring workflow | A repeatable prompt you run every time |
| Vague brief from stakeholder | Forces clarity before you start |
| Team prompt library | Documented, consistent structure anyone can use |
| You keep tweaking the same prompt | One meta-prompt generates better v1s than manual iteration |
What makes a strong task description
The better your input, the better the generated prompt. Include:
- The output format — report, email, list, code, table?
- The audience — your manager, a client, a developer, yourself?
- The success criteria — what makes the result good vs. mediocre?
- The failure mode — what would make it useless or wrong?
Bad: “Write a prompt for summarizing articles.”
Good: “Write a prompt that summarizes technical blog posts into a 3-bullet TL;DR for non-technical readers, without losing the key insight from the author.”
🔁 Leftover Remixes
🌶️ Spicy: After generating the prompt, ask: “What’s the most likely way this prompt produces a bad output? Rewrite the Constraints section to prevent it.”
🧊 Mild: Just ask for the Role and Output Template sections only — that’s often enough to get dramatically better results without the full framework.
💰 Budget: “Rewrite this prompt to be more specific: [paste your existing prompt]” — same structural improvement, less scaffolding.
Inspired by the Meta-Prompt Engine framework. Adapted for practical daily use at The Prompt Kitchen.