6 techniques to fix ChatGPT’s annoying habits

You’ve experienced it. That flash of frustration when ChatGPT, despite its incredible power, responds in a way that feels… off. Maybe it’s overly wordy, excessively apologetic, weirdly cheerful, or stubbornly evasive. While we might jokingly call it an “annoying personality,” it’s not personality at all. It’s a complex mix of training data, safety protocols, and the inherent nature of large language models (LLMs).

You have more control than you think. 

Why does ChatGPT act that way?

Understanding the ‘why’ helps craft better ‘how-to’ prompts. ChatGPT’s quirks often stem from:

• Training data influence: ChatGPT learned from vast amounts of internet text, including forums, articles, books, and websites. It absorbed the patterns, styles, and unfortunately, some of the verbosity and clichés present in that data.
• Reinforcement learning from human feedback (RLHF): Humans rated AI responses during training, teaching it to be helpful, harmless, and honest. This process heavily favoured politeness, clear signaling of its AI nature (“As an AI model…”), and cautious phrasing, which can sometimes lead to excessive hedging or apologies.
• Safety guardrails: To prevent harmful, unethical, or inappropriate output, strict safety protocols are in place. While essential, these can sometimes cause the AI to refuse seemingly innocuous requests or be overly cautious, interpreting prompts in the most risk-averse way.
• Predictive nature: At its core, ChatGPT predicts the most statistically likely next word (or token) based on your prompt and its training. It doesn’t truly “understand” context or nuance like a human, leading to misinterpretations or generic output if the prompt isn’t specific enough.
• Prompt interpretation: How well it performs depends heavily on how clearly it interprets your instructions. Ambiguity leads to unpredictable results.

Common ChatGPT annoyances and how to engineer better responses

Let’s tackle some frequent frustrations with specific prompt engineering techniques:

1. Excessive verbosity

Description: Getting paragraphs when a sentence would suffice; overly elaborate explanations for simple concepts.

Likely cause: Training data often includes detailed explanations; RLHF might favour thoroughness.

The fix: Be explicit about length and format.

• "Explain [topic] concisely."
• "Summarize the key points in 3 bullet points."
• "Answer in a single sentence."
• "Limit your response to under 100 words."
• "Provide a brief overview of [topic]."

Example:

Instead of: “Tell me about photosynthesis.”
Try: "Explain photosynthesis in two sentences suitable for a 5th grader."

2. Constant hedging and apologies

Description: Phrases like “As an AI language model…”, “It’s important to note…”, “I cannot…”, “I apologize for any confusion…” even when unnecessary.

Likely cause: RLHF and safety training emphasizing limitations and politeness.

The fix: Instruct it to be direct and assume user understanding.

• "Answer directly without hedging."
• "Do not apologize or state you are an AI."
• "Provide the information without qualifiers like 'it's important to note'."
• "Assume I understand the limitations of AI models."
• "Be confident in your response." (Use with caution, can increase hallucination risk if topic is complex).

Example:

Instead of: “What are the benefits of Python?”
Try: "List the main benefits of Python for web development. Answer directly, without apologies or stating you're an AI."

3. Unwanted tone

Description: The tone doesn’t match the context – maybe too enthusiastic for a serious topic or too stiff for creative brainstorming.

Likely cause: Trying to maintain a generally helpful and positive persona derived from RLHF; defaulting to a standard tone without specific instruction.

The fix: Explicitly define the desired tone or persona.

• "Adopt a formal and professional tone."
• "Write in a neutral, objective style."
• "Use a casual and friendly tone."
• "Respond with the tone of an expert [field specialist]."
• "Avoid excessive enthusiasm or exclamation points."

Example:

Instead of: “Explain quantum entanglement.”
Try: "Explain quantum entanglement in a neutral, scientific tone suitable for a college student. Avoid analogies that are overly simplistic."

4. Generic or obvious information

Description: Receiving basic, surface-level answers when you need specific details or deeper insights.

Likely cause: Ambiguous prompts; the model defaults to common knowledge found frequently in training data.

The fix: Provide context, specify the desired level of detail, and ask for specifics.

• "Provide specific examples of [concept]."
• "Focus on the [specific aspect] of [topic]."
• "Assume I have foundational knowledge; explain the advanced aspects."
• "Instead of a general overview, discuss the challenges of implementing [technique]."
• "Analyze the pros and cons from the perspective of a [specific role]."

Example:

Instead of: “How to improve website speed?”
Try: "List 5 specific, actionable techniques to improve website loading speed, focusing on image optimization and server response time. Explain the technical implementation briefly for each."

5. Stonewalling or unhelpful refusals

Description: Refusing to answer a seemingly harmless question, often citing safety or limitations.

Likely cause: Safety guardrails interpreting the request as potentially problematic, even if it isn’t; limitations on accessing real-time data or performing certain actions.

The fix: Rephrase, simplify, or focus on underlying principles.

• Rephrase: Ask the question differently, avoiding potential trigger words.
• Break it down: Ask for smaller, less complex parts of the original request.
• Ask for principles: Instead of asking for potentially sensitive specifics, ask for the general rules, concepts, or steps involved. E.g., Instead of “Write code to access X system,” try “Explain the common methods and security considerations for accessing systems like X via API.”
• Check for constraints: Is the request about real-time data (like today’s stock prices) or personal opinions? Acknowledge you know it can’t do those things, but ask for related historical data or common viewpoints.

Example:

If refused: “Generate a marketing plan for a new type of drone.”
Try rephrasing: "Outline the key components of a typical marketing plan for a high-tech consumer product. Include sections like target audience analysis, channel strategy, and budget considerations."

6. Forgetting context or instructions

Description: Ignoring previous parts of the conversation or instructions given earlier in the same chat session.

Likely cause: Limited context window (how much text it can “remember” at once); difficulty tracking complex, multi-turn instructions.

The fix: Reinforce context and instructions periodically.

• Summarize: Briefly restate key context or previous points before asking a new related question. "Given that we previously established X and Y, now explain Z."
• Use explicit references: "Based on the criteria you listed earlier..."
• Custom instructions (if available): Use the Custom Instructions feature to provide persistent background information and output preferences.
• Keep sessions focused: For very complex tasks, consider starting a new chat session to ensure a clean context slate.