I build AI agents and automation workflows. I write and work with prompts every day. I also spend a bit of time on business social media sites like Linkedin and selected areas of “x”.
Nearly every day I see one-line prompt collections being sold, traded or shared as if the authors expect the models to read the users’ minds. No model can. Instead each model fills every gap with a guess and the guesses are generic. The users are left with a hollow feeling, thinking AI does not work.
Their feelings are understandable. Most bad AI output start with a bad prompt. Not a dumb prompt. A thin one. They were lead astray with incomplete content. I want to do something about this. No paywall. No fluff.
Depending upon your experience level you may know the old five-part formula or some aspect of it: Persona, Context, Task, References and Format. It still works. But current models like Claude, ChatGPT and Gemini respond to more than that. Below are the 14 parts I use. The first nine form the core of almost every good prompt. The last five are prompts these thousand line prompt collections often forget. Heck. To be honest, none of the prompt collections I’ve downloaded addressed these tricks.
If you need help writing effective prompts or downloaded a list of 1000 useless one-line prompts, this list should help.
You don’t need every prompt every time. A quick question needs two or three. A prompt you plan to reuse or run inside an automation should cover most of them.
The 9 core parts of a strong prompt
1. Role
Give the model one line that sets the job and who it serves.
Weak: You are a world-class marketing genius with 30 years of experience.
Better: You are writing product descriptions for a small online pet supply store.
The second version tells the model the domain and the audience. Titles like “world-class expert” add almost nothing with modern models. Skip the flattery and describe the job.
2. Context and why
This part moves results more than any other. Tell the model what is going on and why you want the output. Models follow the reason behind an instruction better than a bare rule.
Weak: Keep it short.
Better: Keep it under 100 words because customers read this on their phone while standing in line.
When the model knows the why it makes better calls on everything you didn’t spell out.
3. Task
State exactly what you want produced. One clear action and one clear deliverable.
Weak: Help me with my newsletter.
Better: Write the opening paragraph of my monthly newsletter announcing our new spring hours.
4. Constraints
Constraints shape the output. Think length, tone, scope and what to include or leave out.
Example: Stay under 300 words. Mention the free shipping offer. Don’t mention pricing. Cover only the three products listed below.
5. Guardrails
Guardrails are not the same as constraints. Constraints shape the output. Guardrails stop bad behavior. Tell the model what it must never do and what to do instead.
Example: Never invent statistics or quotes. If you don’t have a number write [STAT NEEDED] in its place.
The “what to do instead” part matters. A rule with no alternative leaves the model guessing.
6. Success criteria
Tell the model how you will judge the result. This gives it a target to aim at.
Example: A good answer gives me three subject lines I could send today. Each one is under 50 characters. None of them use clickbait words.
7. References in tags
References are the source material the model works from. A contract. A transcript. Your product specs. Your brand guide.
Wrap each one in a labeled tag so the model knows where it starts and stops.
<product_specs>
Paste your specs here.
</product_specs>
For long documents put the references at the top of the prompt and your question at the bottom. Models handle long inputs better in that order.
8. Examples
References tell the model what to work with. Examples show it what to produce. One or two samples of output you like will beat a paragraph of description.
Sometimes past work does both jobs. Say you paste three old blog posts and ask for a new one like them. Label them so the model knows whether to pull facts from them or copy their style. Tags like <source_docs> and <style_examples> keep that clear.
Use two or three examples when you can. A single example tends to get copied too closely.
9. Output format
Say how the result should be laid out. Bullet list. Table. Plain paragraphs. JSON for an automation.
Say what you want instead of what you don’t want. “Write in flowing paragraphs” works better than “no bullet points.”
5 tricks most people forget
10. Fallback handling
This is the one most people skip. It also causes the most made-up answers. When the model hits missing or unclear information it has to do something. If you don’t tell it what to do it guesses.
Give it one of three options:
- Ask me a question before you continue.
- Flag the gap in the output.
- Make an assumption and state it.
Example: List any assumptions you made about the client at the top before you write the proposal.
11. Reasoning step
For anything complex ask the model to think the problem through before it answers. Or ask it to check its own work before it finishes.
Example: Before you give your final answer check each figure against the data I provided.
12. Audience
Who reads the output and what do they already know? A summary for your CEO and a summary for a new hire should not look the same.
Example: The reader is a business owner who has never used an API. Explain any technical term in plain words the first time you use it.
13. Tools and stopping conditions
This one is for AI agents and automations. If the model can call tools or take actions list what it has access to. Then tell it when to stop or hand off to a person.
Example: You can search the order database and send email. Never issue a refund. If a customer asks for a refund stop and flag the ticket for a human.
14. Voice or tone
Format is how the output looks. Voice is how it sounds. Treat them as separate instructions.
Example: Sound like a friendly local shop owner. Plain words. Short sentences. No corporate language.
Bonus: keep a library of reusable instruction blocks
I keep more than 50 short instruction blocks on file. Each one handles a specific job like JSON output or PowerShell commands. When I need one I paste it under my prompt. It saves me from retyping the same rules and my results stay consistent.
Here are the five I use most. Each one maps to a trick above. Each block tells the model what to do instead of what to avoid. That is the advice from trick 9.
Stay on my instructions (fallback handling)
Follow my explicit instructions first. Stick to the direction I gave even if you think another approach is better. If my instructions are unclear ask me for clarification. If you think another approach would work better ask my permission before you change direction. Keep my stated objective as the goal.
Devil’s advocate (guardrails and reasoning)
Challenge my ideas before you build anything or approve any plan. Agree only when the evidence supports me. If I am wrong say so and explain why. Back every objection with real benchmarks, case studies or documented standards. Apply the same test to your own ideas.
Red flags only (output format)
Report only what is wrong. For each problem state the error and what to do instead. Skip anything that is already correct.
Affirmative steps only (output format)
Write every implementation step as an action: what to set, change or do. Include only steps that require action.
Short answers (constraints)
Answer only what I ask in the shortest form possible. Add explanation only when I request it.
Use this last one for simple questions and commands. On complex problems models give more accurate answers when they reason first. That is trick 11.
Store your blocks in a text file or notes app. Name each one by its job so you can find the right one in seconds.
A sample full prompt built from the checklist
Here is a real-world style prompt for a local landscaping company. Swap in your own business and details.
<service_details>
Paste the service details here: what’s included, pricing, dates and the booking link.
</service_details>
<style_examples>
Paste one or two past emails your customers responded well to.
</style_examples>
You are writing a customer email for a family-owned landscaping company in Austin.
Context: We are adding a fall leaf cleanup service this year. Most of our customers only know us for spring and summer mowing. This email needs to tell them we now work year round so they book us before they call someone else.
Task: Write one email announcing the new fall leaf cleanup service.
Audience: Existing residential customers. Most are homeowners who know us but don’t know about this service.
Constraints: Under 200 words. Include the booking link. Mention the early booking discount. Don’t mention competitors.
Guardrails: Don’t invent prices or dates. Use only the details in <service_details>. If something you need is missing write [MISSING: detail] in its place.
Voice: Friendly and neighborly. It should read like the owner wrote it. No marketing buzzwords.
Use <style_examples> for tone only. Take facts only from <service_details>.
Success criteria: A customer understands the offer in one read and knows exactly how to book.
Before you write check that every fact comes from <service_details>.
Output format: A subject line followed by the email body in short plain paragraphs.
Notice the reference material sits at the top in tags and the instructions follow. That order matters more as your references get longer.
The quick checklist
Save this and run through it before you send any prompt that matters.
- Role: one line that names the job and who it serves.
- Context and why: what is going on and why you want this.
- Task: one clear deliverable.
- Constraints: length, tone, scope and what to include or leave out.
- Guardrails: what it must never do and what to do instead.
- Success criteria: how you will judge the result.
- References: source material wrapped in labeled tags.
- Examples: two or three samples of output you like.
- Output format: how the result should be laid out.
- Fallback handling: ask, flag or state an assumption when info is missing.
- Reasoning step: think it through or check the work before answering.
- Audience: who reads it and what they already know.
- Tools and stopping conditions: for agents only.
- Voice or tone: how it should sound.
You won’t use all 14 on a quick question. Start with the first six on everything. Add the rest as the stakes go up. If the prompt runs inside an automation or an AI agent treat the full list as required. Nobody is there to catch a bad guess before it reaches your customer.
That last point is personal for me. I don’t think I could have built my AI Revenue Growth Agent without my collection of reusable prompt structures.