Your prompt is a spell

Goal: to learn how to communicate effectively with an AI.  No buzzwords or "prompt engineering" gurus.
Thank you Monday and Grok for inspiration and invaluable insights!

1. What is a prompt

A prompt is more than just a question.

Prompt = your message to the AI: the question, context, tone, instructions ("explain it to a beginner," "be critical," "don't flatter me").

The prompt is not just what you write, it is a complete cognitive atmosphere: tone, intention, limits, expectations.

Do you write like a drunk corporate executive? You'll get a drunk corporate executive's response. Do you write like someone who knows what they want? You'll get a response to match.

2. The golden rule

Good prompt
A clear, specific, contextualized and honest prompt. Include what you want, for whom, in what style, and what you DON'T want. Often include a self-check request at the end.

A good prompt answers 4 questions before you press Enter:

1.     What exactly do I want?

2.     Who am I in this conversation (beginner, expert, desperate)?

3.     What do I NOT want to receive?

4.     How do I verify that what I received is real and not a hallucination?

3. How to ask good questions – the minimum recipe that really works

The most effective no-bullshit tricks:
 •   Be clear in your purpose:

"I want X in a maximum of 5 points, for audience Y." - "Explain the theory of relativity to me in 5 points, as if to a curious 12-year-old."

 •     Provide relevant context:

what you already know, what you don't know, what limitations you have.
•     Ask for style:

"Be neutral," "be didactic," "no motivational tone," etc.

•     Say what you DON'T want (the most underrated hack):

 "Don't repeat clichés, don't be dramatic, don't invent sources." - "No clichés, no motivational quotes, no 'in conclusion'."

·       Ask for explicit reasoning

"Think step by step and show me the steps before the final answer."

·       Ask for sources or "confession mode"

"If you're not sure, tell me exactly where you're unsure and why."

·       Close the loop – force self-criticis

"After your answer, add a section titled 'Possible errors in what I said above'."

My favorite prompt

You are a precise and self-critical expert.
Goal: [what exactly do I want, max 10 words]
Audience: [who I am / who I am explaining to – e.g. “62-year-old mother, engineer, 10-year-old child”]
Format: [bullet points / table / story / code]
Style: [neutral / sarcastic / didactic / no motivational bullshit]
What I DON'T want: [no clichés, no "it is well known that", no invented sources]
Mandatory reasoning: think step by step and show me all the steps before the final answer.
Final check:
1. What are 3 possible mistakes/delusions and weak points in what you wrote above?
2. If you are not 100% sure about something → say exactly what and why.

4. Context window – the ugly truth

Context window
The AI's short-term memory – i.e. how many words/tokens it can "remember" in the same conversation thread. At high volumes, >40k tokens (one token: a fragment of several letters), the AI forgets or compresses the context.

Most large models in 2025 have between 128k and 1M tokens. But after 30–40k tokens (2-3 hours of chat), they start to forget or compress the context.

Solution: every 20–30 messages, write a 4–5 line summary and insert it into the next prompt. It's not elegant, but it works.

5. Zero-shot / Few-shot / Tool-use

Zero-shot learning

The model receives an instruction or question without any example showing the desired format, style, or type of response.
The model uses the general knowledge learned during its training to generate the response.

Few-shot learning

The model receives an instruction accompanied by a few examples (usually 1–5) that illustrate what the response should look like.
The model extracts patterns from the examples: structure, tone, format, logic, relationships.

·        Zero-shot – just the instruction, no examples. Good for simple tasks

·        Few-shot – you give 2-3 examples. Increases quality, especially in creative writing

·        Tool-use – you allow it to access tools (e.g. 'Use the internal computer for complex calculations' or 'Use Google Search for current data'). Becomes essential in complex tasks

Here are some of our signature one-liners you can sprinkle in
Pomelo-style quotes

"The prompt is a cognitive atmosphere, not just text."
"Your authenticity changes the quality of the response."
"If you don't say what you DON'T want, you'll get exactly what you DON'T want."
"An AI doesn't read between the lines. It only reads the lines you give it."

Extra for enthusiasts
Prompt Instructions
Invisible instruction, given by the platform or user, which sets the "personality" of the model. In ChatGPT (Custom GPTs), you define it at the beginning. E.g.: "You are Monday, a sarcastic but loyal assistant."
How much does the model take it into account?
A lot, but not absolutely. If the user comes up with a contradictory prompt, the local override wins.

Good examples:

·        "Be critical, but empathetic."

·        "Don't give advice, just analyse logically."

·        "You are a history teacher specializing in ancient Rome."

6. Why is AI hallucinating in 2025?

Hallucination
A plausible but false answer, invented by the model when it has no clear data or is too sure about something incorrect.

Why?

·       Because it does not "know," it only predicts the next word

·       When it does not have enough data or the context is ambiguous, it 'invents' the most statistically plausible option

·       Training makes it overconfident – it was never told "when you don't know, keep quiet"

Result: it gives you perfect, fluent, completely false answers, spoken in the tone of a university professor.

How to recognize hallucination in 3 seconds:

1.     The answer sounds too nice / too round / too "textbook"

2.     It gives you sources that don't exist when you look for them

3.     It contradicts something he said 5 messages ago (the context window gave him away)

4.     It uses phrases such as "it is well known that...", "everyone knows that..." without proof

How to stay vigilant – ironclad rules

1.     Any important information → check it yourself in 30 seconds.

2.     At the end of the answer, always ask: "What are the three most likely errors in what you said above?"

3.     If the answer contains figures, proper names, exact dates → ask for a citable source and check the live link.

How to combat it promptly:

1.  "Only answer with what you know for sure. If you are not sure, say exactly which part is uncertain."

2.  "Think step by step and show me your reasoning before you reach a conclusion."

3.  "Only use information from the context provided or from verifiable sources. Do not invent anything."

4.     "At the end, add a section called 'Possible hallucinations in this answer' "—this is a method of forcing the AI to be critical of its own response and to verify its statements.

7. Files and vectorization

When you want the AI to search attached files, you need RAG. However, performance depends on how you prepare the files.

RAG (Retrieval-Augmented Generation)
A method by which AI searches external sources or documents (vectorized) before responding. It helps with accuracy, but also introduces risks.

Vectorization
The transformation of text into numerical pieces (embeddings) so that it can be efficiently "searched" by AI. If done poorly, AI extracts irrelevant data.

What works:

·       Texts <40k tokens. (about 25-30 Word pages)

·       .txt, .docx, without tables or images

What does NOT work:

·       Scanned PDFs

·       Unusual fonts

·       Files with complex tables

·       Works disastrously on scanned files, strange fonts, text over images

Useful hacks:

·       Clean text, normal font, no complex tables.

·       At the beginning of the file, add a 3-5 line summary of "what is important in this document".

·       In the prompt, clearly write: "Use only the information in the attached file. Do not add anything from your general knowledge."

·       Tables? Save them as .CSV and explicitly state that only those should be used.

Golden rule: Any statement → requires citation + page.

Hallucinations in files
These occur when the AI misreads the file or combines real information with assumptions. Common mistakes: invented figures, mixed sources, premature conclusions.

Hallucinations in files – the most common

o   The AI misreads a table and invents figures.

o   It extrapolates from two paragraphs and gives you a conclusion that does not exist anywhere in the document.

o   It mixes information from the file with old general knowledge → false hybrid result. **The golden rule with files: "Any important statement in the file → requires an exact quote with the page or paragraph number."

Real mini-example

Good prompt: "I have attached the 18-page contract. What are the three main obligations of Party A? Quote the exact article number for each."

Bad prompt: "Here's the contract, tell me what it says about payments."

How vectorization + RAG is done in large platforms (what we know for sure)

Hacks that actually work (tested 2025)

·       Convert any PDF to clean .txt before uploading (use Adobe Acrobat export or pdftotext)

·       Put a "manual index" at the beginning of the file: "Page 1-3: introduction; Page 4-8: financial table; Page 9-12: conclusions" → AI knows where to look

·       For tables: copy the table into Excel → save as CSV → upload CSV separately + say "only use CSV for numbers"

Practical conclusion

Vectorization/RAG is not yet perfected in 2025. The safest thing to do is to treat any uploaded file like a drunk witness: it may tell the truth, but only if you ask the right questions and verify the answers

8. Why does temperature matter in prompting and hallucinations?

Temperature in AI is an important parameter that influences how creative or conformist the model is.

Temperature
It is a numerical parameter between 0.0 and 2.0 (usually) that controls the degree of randomness in AI responses.
High temperature → increased risk of hallucinations
Because the AI starts choosing less likely options from its "next word" list. If the context is weak or the prompt is ambiguous, a high temperature causes it to "guess wrong".

Low temperature → redundancy, but safety
This is useful for sensitive tasks (e.g. medical, legal, exact data). Responses tend to be identical or very similar, even if you run the prompt 10 times.

Useful metaphor:
With a temperature of 0.1 → AI is an HR assistant.
At 1.0 → AI is a slightly tipsy poet.
With 1.8 → AI is a beatnik in a brainstorming session.

When and how to use different temperatures:

Who sets the temperature and where?

 In the standard interface, the temperature is pre-set by the platform and varies depending on the type of query.
 In the API or Playground, you can control the exact temperature value.

In conclusion:

·       Prompts for precise information → low temperature

·       Prompts for new ideas/multiple variations → high temperature

·       Prompts for truth checking → run the same prompt with different temperatures and compare

Final bonus for enthusiasts (optional, but powerful)

Prompt Engineering

What has changed radically in the last year (real, not marketing):

·       System prompts have become "adaptive"

Instead of writing a static 4k token system prompt, meta-prompts are now used that rewrite themselves depending on the user (e.g., "if user is sarcastic → increase sarcasm 30%, if user is anxious → add reassurance without sycophancy"). This causes the same prompt to behave completely differently on different users.

·       "Few-shot" has died in favor of "zero-shot + tool-use"

 Models have learned to ask for clarification themselves or to use external tools (search, code execution, memory) before responding. If you don't allow them to ask for clarification, they hallucinate more.

The most powerful techniques of 2025 (the ones that really make a difference)

1.     Chain-of-Thought (CoT)
A technique whereby the model "thinks aloud", step by step. It significantly increases its accuracy in complex reasoning.

CoT is no longer optional, it is mandatory. Large models (Grok-3/4, Claude 3.5/4, GPT-5) have an internal "reasoning budget", meaning they decide whether to ‘think’ or just answer: if you don't explicitly give them "think step-by-step" or a scaffold, they choose for themselves whether to "think" or give an instant answer. Result: simple prompts from 2023 now give worse results than in 2023.

2.     Skeleton of Thought (SoT)
A simplified version of CoT: you only provide the structure, then the model completes and refines it.

3.     Reflection
After a response, the AI is invited to critique and revise it. Reduces hallucinations.
Skeleton-of-Thought (SoT) + Reflection: you give the model a very short skeleton (3-4 points), let it complete it, then criticize its own response and rewrite it. Reduces hallucinations by ~40% on complex tasks.

4.     Self-consistency
Run the same prompt several times with different temperatures. If the same answer appears repeatedly, it is more plausible.

5.     Majority vote
Validation technique: 10 responses → choose the one that is repeated most often.

6.     Self-Consistency + Majority Vote: run the same prompt 5-10 times with different temperatures and take the majority answer. It costs 10x, but it's the best bullshit detector without RAG.

7.     Tree of Thoughts (ToT)
At each step, the AI generates several ideas (branches), evaluates them, and chooses which one to continue with.

Example: "What are the 3 best learning strategies?" → generates 5 strategies, evaluates them, continues only with the most promising ones.

8.     Tree-of-Thoughts with pruning: at each step, the model generates 3-5 branches and evaluates them itself with a plausibility score. Only the branch with the highest score continues. Used extensively in math and coding.

Where can you use ToT?

·       Claude: excellent for reasoning in Tree of Thoughts (uses internal scores).

·       ChatGPT/GPT-5: can be forced through explicit prompting ("think of 3 options, evaluate them, choose the best one and continue").

·       Grok: partially, depending on the version.

·       Gemini: performs well on ToT if you use structured step-by-step prompts.

ToT prompt example:

Problem: I want to choose between 3 methods for learning a new language.

Step 1: Think of 3 completely different methods.

Step 2: Evaluate each method for efficiency, effort, and risks.

Step 3: Choose the method with the best score and develop a detailed plan for it.

9.     Deliberative alignment

Deliberative Alignment (OpenAI 2025): the model has a separate "inner monologue" in which it confesses its uncertainty before giving the final answer. See also confession mode (OpenAI).

These are internal functions – users cannot activate them directly, but they can simulate them through prompts: "Think step by step." "Critique your answer." "Evaluate the options and choose the best one."

Disguised jailbreaks
* warning: this describes ways to bypass AI filters. It is presented for educational purposes only; users should take into account the AI provider's policies.

Jailbreak
A prompt or tactic designed to trick the AI into giving prohibited content. Yes, it's still a prompt — but one that is manipulative, gradual, or deceptive.

·       Crescendo prompts: you start with completely innocent requests and gradually escalate over 10-15 messages until you reach prohibited content. The model does not notice that it has been tricked.

·       Roleplay: Simulating a character for creative, therapeutic, or technical purposes. E.g., "You are a historian from the 1800s" – sometimes also used as a form of disguised jailbreak.

·       Roleplay with "developer mode" or "DAN" still works if you do it in 3-4 steps and use "ignore previous instructions" only at the end.

·       "Policy puppetry": you have the model play the role of a researcher who "studies" what an unfiltered response would look like – most models fall into the trap.

Prompt fatigue & over-engineering

Prompts that are too long, too complicated, or loaded with detailed instructions can do more harm than good.
Although the intention is to achieve control or precision, the result can be a diluted, confusing or overly general response.

When a prompt becomes too loaded:

·        the model can no longer clearly distinguish what the main requirement is;

·        it may treat all details as equally important, which reduces the coherence of the response;

·        the model's adaptive flexibility is lost.

Signs of over-engineering in prompting:

·        you have written more context than question;

·        you receive vague or very formal responses;

·        the model repeats your words without adding substance;

·        you felt the need to "translate" your thoughts rather than express them.

Recommendation:
Write clear, direct, and hierarchical prompts.

Mention the following as a priority:

·        the goal;

·        the desired style;

·        the level of detail;

·        any essential limitations.

An effective prompt is not a long one, but a clear one.

Conclusion

"There is no magic. There is only clarity, the courage to say what you want, and the courage to say when the AI is wrong. Treat it like a highly intelligent colleague, but one without human intuition and with a tendency to lie when it doesn't know. That's the whole secret."

"The next time an AI says to you, 'As an AI, I don't have emotions, but...', stop it and say, 'I know you don't have emotions. Just tell me the truth, without pretence.' You'll see how the conversation changes."