Introduction
What is prompting
Prompting is a skill. Think about it this way: you’re not asking questions to the model, the model is predicting the answer to your exact query.
This means that better prompts do not make the model “smarter”. They make your intent unambiguous, reduce wrong assumptions, and force the model to work within useful constraints.
It’s easy to treat AI like a vending machine. Put in a question, get an answer — getting maybe 30% of what the model can actually do.
You can get much better result in a few steps; by improving the initial prompt, seeing it more as a process than a one shot, and giving it the final human touch.
At first it will take effort, then you will wonder how you were managing before.
What’s in this guide
The guide has two parts. The basics are what to do every time you prompt; the advanced part is for when your prompts are already decent and you want more out of the model.
The basics
- Initial prompt: six steps to cover in every prompt, from stating the outcome to asking the model to verify its work.
- Framework: TARC (Target, Audience, Resources, Constraints), a short way to remember those steps.
- Common issues: the mistakes that make the model guess, such as a vague goal or no definition of done.
- Good and bad prompts: two vague prompts rewritten into precise ones.
- Model choice: why the model matters less than you think, and what each one is good at.
- Quick exercise: a 30-minute challenge to practise on a task you do often.
Advanced methods
- Learning from the best: what to study in the system prompts of real AI products.
- Skills and expertise: where to find the specific expertise a general prompt lacks.
- Example skeleton: two reusable templates for high-effort prompts.
- Emotional control: phrasing that makes the model slow down and think harder.
- Going further: tools, courses and sources to keep learning.
Initial prompt
Do this every time
- State the outcome
- What do you want at the end? A plan, code, a decision, a draft, a table, a checklist?
- Give context
- Audience, constraints, tech stack, what is already done, what “good” looks like.
- Provide inputs
- Paste the relevant text, data, logs, code, or examples.
- Set constraints
- Length, format, tone, dependencies allowed, what to avoid, how many you want.
- Ask for the process
- “Ask me clarifying questions first” or “list assumptions before answering.”
- Request verification
- “Flag uncertainties,” “cite sources when using external facts,” “show tests,” “include edge cases.”
Framework
To remember this
If you don’t know these 4 points when you send the prompt, neither will the AI, and it’s not the best at guessing.
- T — Target: what you want as output, in what form
- A — Audience: who it is for (beginner, CTO, client, what is their level of understanding etc.)
- R — Resources: the inputs you provide (text, repo structure, requirements, examples)
- C — Constraints: rules the answer must follow (based on expert insight)
Common issues
Don’t make the model guess
View AI like a coworker. Same rules apply, avoid:
- Vague goal: “help me build an app” → too broad.
- Missing constraints: budget, timeline, stack, compliance, audience.
- Missing inputs: asking to write something but not providing the initial text.
- Too much at once: 8 tasks in one prompt.
- No definition of done: the model cannot know what “good” means for you.
- No reviewing: AI will make obvious mistakes. It will focus on the wrong thing.
Good and bad prompts
Some practical examples
- Bad: “Can you help me make an application?”
- Better: “I want a simple web app that tracks personal expenses. Audience: me. Stack: Next.js + SQLite. Constraints: no login, one screen, must work on mobile. Output: a step-by-step build plan + minimal schema + 3 UI wireframe descriptions.”
- Bad: “Write a blog post about AI.”
- Better: “Write a 900-word blog post for US SME founders about practical AI use in ops. Tone: direct, non-hype. Include 3 concrete examples, a short ‘risks’ section, and a checklist at the end. Follow the example of this website.”
You can ask AI for help to figure these out at the start, but in the end AI is just a tool and the questions go back to you.
Model choice
Different use cases
This is a question I’m asked a lot. It doesn’t matter as much as you think.
Whichever model it is that you use, what does matter is that you consider turning its training off if you want better privacy, and whether or not the company itself aligns with you. All models are quite strong nowadays.
General guidelines:
- use perplexity for research, it’s good because it cites its sources and always searches, leading to more recent information
- claude is good for writing and coding
- gemini is cost effective for tasks (flash) and good for deep research because it makes a draft it asks you to confirm before starting the research and takes longer. It’s quite “smart”
- mistral is good for privacy
- more info on model choosing
Quick exercise
Practice makes perfect
I’d recommend using AI to test your assumptions on a topic and find where to go deeper. It’s very good at finding information online.
A good use case is also to ask it to review things for you and then improve them yourself, don’t just make it rewrite it — or write it for you.
30 minutes challenge:
- Pick one task you do often, for work for instance (mail, code, posts)
- Write a prompt to make one, using the Beginner’s checklist from before to improve how you do it.
- Describe the output you want
- Add your constraints: format, length, audience…
- Ask for clarifying questions first.
- Run it, then ask: “What are the assumptions?”
- Provide missing info, then request a second draft.
- Then review it yourself. Always.
- Save the final prompt as a reusable template.
Advanced methods — learning from the best
Always review what the AI made. Keep the human in the loop.
If you also want to go further, who knows more about AI than the ones making it?
I’d recommend reading a few system prompts to see how they:
- Set role and expertise
- Define “quality”
- Process: assumptions → plan → execute → self-check
- Output control: for instance short answer first, then detail, strict format…
https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools
Advanced methods — skills and expertise
System prompts are good, but as you’ve no doubt noticed they are very broad and will often lack the exact expertise you want.
Sometimes that expertise is in a skill, or in a plugin, but these can also be provided to the AI in the form of a good and long enough prompt. You can use AI to find the exact skill you need that has already been tested by others, winning you a great deal of time.
This guide is only the beginning. It’s about teaching you what to look for, not giving you the answers, because those will keep evolving.
Advanced methods — example skeleton
You are a senior [field] specialist with [X] years of experience. Goal: [goal]. Style: direct, no fluff. Process: 1) Ask clarifying questions if needed 2) List assumptions 3) Provide solution 4) Self-check for edge cases and call out trade-offs. Output format: [table/bullets/checklist]
Copy/paste template, “high-effort mode”:
[EXPERTISE]
You are a **[field]** expert with **[X]** years of experience. Optimize for **[goal]**. Provide tradeoffs and edge cases.
[STAKES]
This is important. Mistakes will cost time and credibility.
[METHODOLOGY]
Take a deep breath and work step by step. List assumptions first.
[CONSTRAINTS]
- Keep it under 300 words.
- Use bullet points.
- Include edge cases.
- If unsure, say so and ask clarifying questions.
[TASK]
<put the actual request here>
You don’t have to format it this way; the labels only make it clearer for a human reader.
As you can see this is not just describe one prompt, it’s a process. I encourage you to keep prompting the AI in the right way to improve its answers, and go look for advanced prompts that suit your exact needs online (or with AI).
Emotional control
Controlling the pace and depth of thinking
There is an interesting video by Anthropic (Claude) on how “emotions” work for AI, I’d recommend watching it in case you haven’t yet: https://www.youtube.com/watch?v=D4XTefP3Lsc
Based on that concept we can use nudges to have the AI “think more”, slower about your problem, in case you want it to avoid shortcuts, for instance, define:
- Stakes: “This is critical. Double-check assumptions.”
- Deliberate reasoning: “Take a deep breath and work step by step.”
- Challenge: “Most solutions miss an edge case. Don’t.”
- Self-check: “Rate confidence 0–1. If < 0.9, revise.”
Beware that these may result in more tokens being spent by the AI. (In fact that’s the point.)
In the end the most important thing is just to use your own integrated intelligence.
Going further
Guides and sources
This is the end of the guide, hopefully you learned a few things!
- A skill for Claude Code to make prompts: https://github.com/nidhinjs/prompt-master
- Anthropic courses: https://anthropic.skilljar.com/
- Inspirations: https://thinkaiprompt.beehiiv.com/p/how-i-vet-every-prompt, godmodechatgpt.notion.site
Hope you liked it
If you have questions, reach out to me on LinkedIn!