Prompting for Accuracy: Engineering Logic

Most hallucinations can be "Prompted Away." By changing how the model processes its thoughts, we can force it to be more careful and precise.

1. COT (Chain-of-Thought)

Don't let the model jump to the answer. Force it to show its work.

• Prompt: "Think step-by-step. First, identify the user's intent. Second, list the tools needed. Third, identify any missing information."
• Result: Because the model "writes down" the logic, it catches its own mistakes before it reached the final output.

---

2. Few-Shot Prompting (The Gold Standard)

The most effective way to eliminate hallucinations is to give "Examples" of what a good answer looks like.

• Provide 3-5 pairs of Question -> Correct Tool Call -> Correct Answer.
• The model will copy the Logic Pattern of the examples.

---

3. Negative Constraints

Tell the model what NOT to do.

• "NEVER assume a user's location. If not provided, ask the user."
• "DO NOT use the calculator for numbers under 100." Models are often better at following "Bans" than they are at following generic goals.

---

4. Visualizing the Prompt Structure

[SYSTEM PROMPT]
You are a reliable analyst.
Rules:
1. Always cite sources.
2. If uncertain, say "I don't know."

[EXAMPLES]
Q: Who is the CEO of Nvidia?
Thought: Search for Nvidia CEO 2024.
Answer: Jensen Huang.

[USER QUERY]
Q: Who is the CEO of Apple?

---

5. The "I Don't Know" Directive

By default, an LLM feels "forced" to answer. You must explicitly give it permission to fail.

• Instruction: "If the search results do not contain the specific answer, strictly state: 'I am unable to find this information.' Do not guess."

---

Key Takeaways

• Step-by-step thinking drastically reduces logical jumps.
• Examples (Few-Shot) are more powerful than 1,000 words of instructions.
• Permissions to fail prevent the "Desperation Hallucination" where the model lies to be helpful.
• Specific citations turn the agent from a "Guesser" into a "Reporter."