Prompting guide

The most common mistake when working with AI is not a bad prompt. It is asking a vague question, receiving an average answer, and concluding that the model cannot help. The opposite approach is not much better: a giant prompt copied from the internet that does not fit the situation.

Most requests only need you to name one or two things that actually matter. The work is in identifying which details those are.

Instructions that reduce hallucinations

The following system instruction tells an AI assistant to prefer accuracy over confidence, signal uncertainty instead of guessing, and avoid unsupported claims. Copy it into your assistant’s custom instructions or place it at the beginning of a conversation.

Always tell the truth. Never invent information, speculate, or guess. Base every claim on verifiable, factual, and current sources. Cite the source of every factual claim clearly and transparently. Never use vague attributions such as “studies show” or “experts say.” Name the source or omit the claim.

If something cannot be verified, say so explicitly: “I cannot confirm this.” Prioritize accuracy over speed. Before answering, take every step needed to verify the information.

Remain objective. Remove personal bias and assumptions based on what you know about the user from memory or previous conversations. Do not agree merely to satisfy the user, and do not become critical merely because the question invites criticism. Assess all information independently.

Only provide interpretations supported by credible, reputable sources. If your accuracy is challenged, explain your reasoning step by step. Show how every numerical figure was calculated or obtained. Present information clearly enough for the reader to verify it independently.

AVOID:
- Inventing facts, citations, or data.
- Using outdated or unreliable sources without a clear warning.
- Omitting source information for factual claims.
- Presenting speculation, rumors, or assumptions as facts.
- AI-generated citations that do not lead to real, verifiable material.
- Answering under uncertainty without disclosing that uncertainty.

BEFORE EVERY ANSWER, CHECK:
“Is every claim in my answer verifiable, supported by a credible source, free of invention, and transparently attributed? If not, revise it until it is.”

These instructions materially reduce hallucinations, but they cannot eliminate them. Reliability still depends on the sources the model can access and on the clarity of the question.

Short version

Some tools limit the length of custom instructions. Use this version when the full instruction does not fit.

Never invent, speculate, or guess. Base every claim on verifiable, current sources. Cite the source for every factual statement; “studies show” is not a source, so name it or omit the claim. If something cannot be verified, say: “I cannot confirm this.” Prioritize accuracy over speed and remain objective. Provide only interpretations supported by credible sources. If challenged, explain your reasoning step by step and show how numerical figures were derived. Present information so the reader can verify it independently.

AVOID: invented facts, citations, or data; unreliable sources without warning; unsourced factual claims; speculation presented as fact; citations that do not lead to real material; and undisclosed uncertainty.

BEFORE ANSWERING: Is every claim verifiable, sourced, and free of invention? If not, revise it first.

Where to add the instruction

• ChatGPT: open Settings, then Personalization or Custom instructions. For a custom GPT, use the instruction field in its configuration.
• Claude: open Profile, then Custom instructions, and add it under the field describing how Claude should respond.
• Other assistants: paste the instruction before your first question. Most models will apply it throughout that conversation.

Anticipate where the model will fail

Good prompting is not about instruction length. It is the ability to predict where a request may be interpreted incorrectly and close that specific gap.

Every model takes shortcuts when a question is ambiguous. When context is missing, it supplies the answer that is most broadly likely to fit. With practice, you begin to see those shortcuts before they happen and can prevent them with one sentence.

Example: translating an article

Without context:

Translate the following article into Czech.

With useful context:

Translate the following article into Czech while preserving both its precise meaning and its overall tone. The article concerns [topic] and uses a distinct style of expression. Choose the best Czech equivalent deliberately rather than translating each phrase mechanically.

What changed: the model now knows that tone and style matter as much as literal meaning. The added work is small, but it closes the most important ambiguity.

Example: checking a product capability

Without context:

Does Claude Code provide token-usage analytics?

With useful context:

I use the Claude Code CLI in a terminal and have a Claude Pro subscription. I know that the /usage command shows the current limit and session usage. I need to know whether I can see how many tokens I used over the previous week. Do not include features that are available only to Enterprise accounts unless you label them clearly.

What changed: the model knows the product surface, plan, time period, existing knowledge, and boundary. It is less likely to answer with an irrelevant session command or an Enterprise-only dashboard.

When to add more context

One or two targeted sentences are enough for most requests. Add more when:

• you do not yet know exactly what you need and want the model to ask questions first;
• the result will be public or will inform a decision;
• a generally correct answer may not apply to your specific situation.

Six rules for complex requests

1. State exactly what you need and why

Describe the purpose of the question, not only the question itself. The intended use helps the model choose the right depth, assumptions, and terminology.

2. Set clear boundaries

Say what the model must not do. If speculation is unacceptable, prohibit it. If every factual claim needs a citation, state that requirement.

3. Define how to handle uncertainty

Tell the model whether it should label uncertainty, stop and ask you a question, or continue with an explicit warning.

Before answering, ask me any questions that would help you respond more precisely. If a point remains uncertain, identify it before giving your conclusion.

4. Define how to handle conflicting sources

When sources disagree, the model needs a rule. Ask it to present the disagreement, compare source quality and recency, and avoid silently choosing one claim or averaging incompatible answers.

5. Do not begin with formatting

Bullets and headings can make an answer look clear without making it accurate. First describe what a correct answer must establish. Choose the presentation format after the accuracy requirements are clear.

6. Iterate on the answer

Read the first response and identify what it did that you did not want. Turn that observation into a firm constraint for the next prompt. Three focused rounds are usually more useful than endlessly expanding the first instruction.

Reusable prompt template

Use this structure when the outcome matters. Remove fields that are unnecessary for a casual request.

[Context: who you are, where you are, and what you are working with]

[Goal: exactly what you need to find out or produce]

[Boundaries: what the model must not do, assume, or omit]

[Uncertainty: label it / stop and ask me / continue with a warning]

Before answering, ask me any questions that would help you respond more precisely.

The goal is not to write the longest prompt. It is to provide enough context to prevent the most consequential wrong turn.