6. Lesser-known but high-impact prompt engineering tips

By now you are equipped with the fundamentals and plenty of examples. In this section, we’ll explore some advanced tips and tricks in prompt engineering that aren’t as commonly discussed, but can significantly improve the quality of your AI interactions. These tips will help you handle more complex tasks and edge cases where basic prompting might not suffice. Even if some of these seem technical, they’re worth understanding as a power user of AI tools.

• Use delimiters to clearly separate sections: When your prompt involves providing data or long text for the AI to act on (like a chunk of user feedback or an excerpt of a report), it’s wise to delimit that section clearly. Common delimiters include triple backticks for large text, quotes, or markers like <<DATA>> before and after the data.

  For example: Prompt: “Summarize the key points from the following user feedback: [paste feedback here] ```.” The triple backticks signal clearly what text is to be summarized. This prevents the AI from confusing your instructions with the content.

  Another way is: “Feedback starts here << *** >> [feedback] << *** >>. Summarize the feedback focusing on complaints.”

• Let the AI “think” with longer outputs or chain-of-thought: Sometimes, if you ask for a very short answer, the AI might not do enough reasoning. There’s an insight from AI research that LLMs kind of “think by writing” – they have no hidden scratch pad; the generation of output is their way of reasoning. So if you force an extremely brief response, you might get a shallow answer. One technique is to allow or even request a longer reasoning first, then the final answer. For instance, for a complex problem you could prompt: “Explain step by step how you arrive at the solution, then give the final recommendation in one sentence.” This makes the model show its work. Another approach is multi-turn: first prompt, “Think through the possible solutions to X in detail,” then after it responds, you follow up with “Great, now given that, what’s your final recommendation?” This two-step approach (first reason, then answer) can yield more robust results. In fact, an expert suggests explicitly doing Reason then Format as two phases. You reason in one prompt (possibly a verbose answer), then in a second prompt ask it to format or summarize that reasoning into the desired output. This is a way to tackle very complex tasks by breaking them down.

• Chaining prompts (sequential prompting): Building on the above, don’t hesitate to use multiple prompts to achieve one goal. You can think of it as intermediate steps – just like you wouldn’t solve a big problem in one leap, you can guide the AI stepwise. For example, you have a raw user feedback dataset. Step 1: Prompt to extract key complaints Step 2: Take those complaints, prompt to brainstorm features to address them Step 3: Prompt to write a proposal for the top feature. Each step’s output is used in the next prompt. There are even tools and libraries (for developers) to chain prompts programmatically. But as an end user, you can do it manually in a chat. One caution: with more steps, there’s more room for error or drifting context, so keep it as tight as needed – often 2 steps is enough, as more can become unpredictable. The “chain of thought” methodology has proven effective: an AI might not get it perfect in one go, but if you correct or refine the question in steps, you converge to a good answer.

• Output primers and format hints: We touched on this earlier – the idea of priming the output by ending your prompt strategically. Let’s say you want the answer in a bullet list. You could end your prompt with something like: “List:\n1.” The AI will likely continue with point 2, 3, etc. Another example: If you want a JSON output (for more technical tasks or to feed into something else), you can end the prompt with an opening curly brace { and maybe a key, like {“result”: … The model often will continue in JSON form. Or simply say “Output only valid JSON.” These hints dramatically narrow the model’s response pattern.

• No need to be overly polite or verbose: People often start with “Please” and long-winded niceties to an AI. It doesn’t hurt, but it uses up context and sometimes muddles the instruction. LLMs do not require polite phrasing – they’re not truly sentient to appreciate it (though ensure you’re not rude either, because that can confuse intent). For instance, “Could you maybe sort of give me some ideas about marketing, please?” is less clear than “Give 5 marketing ideas targeting college students.” One prompting expert explicitly says there’s no need to be polite with LLMs. Direct and clear is the name of the game. Think of it like talking to a search engine or a robot that has no feelings – be concise and straightforward. Similarly, avoid fluff like “I was wondering if you could help me with the following problem I have…” – just state the problem. The AI won’t get offended by a terse command; in fact, it usually prefers it.

• Experimentation is key – iterate quickly: Even prompt masters rarely hit the bullseye with every first prompt on a complex task. Don’t be afraid to run multiple experiments. The feedback loop is quick: you try a phrasing, see the result, adjust. This iterative approach is how many great prompts are discovered. For example, if the output is too generic, maybe you add more context or constraints and try again. If it’s off-track, you might realize your question was ambiguous and reword it. Treat it a bit like A/B testing. There are even emerging tools that let you save and test prompts systematically. You don’t necessarily need a special tool – just the willingness to tweak and run again. The beauty with ChatGPT and similar interfaces is you can refine it by saying “Actually, ignore that, instead do this…” or just start a new chat with adjusted wording.

• Try different models or tools for different strengths: All AI tools are not identical. ChatGPT (especially GPT-4) is great all-around, but there are also others. Some may handle certain tasks better or have larger context windows (meaning they can take in more text at once). If one model isn’t giving what you need, there’s no harm in trying the prompt on another. For instance, some users find that for summarizing very long texts, a model like Claude might do better. Others find differences in creativity or style. At the time of writing, GPT-4 is one of the most capable for complex reasoning, but it’s also slightly slower and has lower availability in some cases. Being a savvy PM, you can leverage the right tool for the job. This is akin to how you might choose the right spreadsheet or the right analysis method depending on the need. The underlying tip here is don’t assume the AI is one monolithic thing; sometimes switching contexts or versions can yield a difference.

• Use the AI to improve your prompts: This one is a bit meta, but you can actually ask the AI how to prompt it. For example, you can literally input: “How should I prompt you to get a detailed competitive analysis of X? What information do you need from me?” The AI might respond with something like, “Provide me with context on X, specify which competitors, and what aspects to compare.” Or if an output isn’t great, you can share it and ask, “How can I improve my request to get a better answer?” This can give insight into what the model “expects”. It’s a lesser-known trick to basically use the AI as a tutor for prompt engineering itself. That said, take its advice with a grain of salt – sometimes it might suggest overly complex prompts. But often it will remind you to add specifics or clarity.

• Stay updated on model behavior changes: Models like ChatGPT can update over time, which might affect prompting quirks. A prompt that once worked might suddenly behave a bit differently if the model got a new training cut or changes in formatting. The field is moving fast, and best practices in 2024 might shift by 2025 slightly. For example, early on people found “let’s think step by step” made GPT-4 give better answers for tricky problems (it induced a chain-of-thought). As models improved, they sometimes do that reasoning without being told, but in some instances you still might want to explicitly request it. So, if you catch wind in the AI community that a new technique or model version is out, be ready to adapt. The good news is the fundamentals we’ve covered remain fairly constant (clarity, context, etc.), but advanced techniques might come and go.

By experimenting with these advanced techniques, you’ll handle edge cases much better. For example, suppose you have a very nuanced task – maybe generating a pseudo-code from requirements then explaining it. A chain of prompts with role-play (first have it be a coder, then a teacher explaining) might do wonders. Or if you get jumbled outputs when providing tables, using clear delimiters or asking for a specific format like CSV can fix that.

These are the kinds of tricks that prompt enthusiasts use to push the limits of what AI can do, treating the AI almost like a very flexible but somewhat literal-minded colleague: you sometimes have to walk them through the process (chain prompting) or give them a gentle push (output primer) to get the best results.

Don’t be overwhelmed by these – you don’t need them for every task. But keep them in your back pocket. Next time you’re slightly disappointed with an answer, think, “hmm, what if I break this into steps?” or “maybe I should give it an example.” With practice, this becomes an intuitive part of prompt engineering.