Patrick José Pereira - ai

Patrick José Pereira - ai https://patrickelectric.work I am an Electronics Engineer that works with robotic systems, OSS and OSH. Zola en Wed, 03 Dec 2025 12:45:51 +0000 How to rock: The Ultimate Guide to Creating Better AI Prompts Wed, 03 Dec 2025 12:45:51 +0000 https://patrickelectric.work/blog/2025/ultimate-guide-to-creating-better-ai-prompts/ https://patrickelectric.work/blog/2025/ultimate-guide-to-creating-better-ai-prompts/ <p>After a couple of nights fighting with LLMs, I noticed that most "bad outputs" were actually <strong>my</strong> lack of clarity. This post is a brain dump so you can avoid my detours and ship prompts that behave.</p> <span id="continue-reading"></span> <p>TLDR: personas + context + structure beat "hey, write this for me" every single time.</p> <h1 id="the-core-principle">The core principle</h1> <p><strong>AI prompting is programming with words, not asking questions.</strong> You're not chatting, you are instructing a probabilistic completion engine. Treat prompts like mini specs: inputs, constraints, outputs.</p> <p>If you can't explain what you want to yourself, the model will improvise (is this a feature or a bug ? ¯\_(ツ)_/¯). All prompting problems are thinking problems. Write the gist in your notebook, then paste it.</p> <h2 id="essential-techniques">Essential techniques</h2> <ul> <li><strong>Use personas</strong> (<a href="https://en.wikipedia.org/wiki/Persona_(psychology)">ref</a>) <ul> <li>Tell the AI who it is and who it's talking to. This helps cut the noise.</li> </ul> </li> <li><strong>Provide context (ABC: Always Be Contexting)</strong> (<a href="https://medium.com/@unstitution21/why-context-is-everything-a1db99b28671">ref</a>) <ul> <li>Include all relevant facts, even the obvious ones.</li> <li>Give explicit permission for the AI to say "I don't know", this reduces hallucinations.</li> <li>Remember: any missing context, the AI will fill in with guesses.</li> </ul> </li> <li><strong>Define output requirements</strong> <ul> <li>Specify the format, tone, and any other constraints for the response.</li> </ul> </li> <li><strong>Use few-shot examples</strong> <ul> <li>Show 2–3 written samples in the style or format you want. The more concrete, the better.</li> </ul> </li> </ul> <h2 id="advanced-techniques">Advanced techniques</h2> <ul> <li><strong>Chain of Thought (COT)</strong> (<a href="https://arxiv.org/pdf/2201.11903">ref</a>) <ul> <li>Ask the model to show its reasoning steps before writing the final answer.</li> </ul> </li> <li><strong>Tree of Thought (TOT)</strong> (<a href="https://arxiv.org/pdf/2305.08291">ref</a>) <ul> <li>Request multiple solution options, then ask for a synthesis or best-of.</li> </ul> </li> <li><strong>Battle of the bots (adversarial validation)</strong> (<a href="https://arxiv.org/pdf/2410.04663">ref:D3</a>, <a href="https://www.mdpi.com/2076-3417/15/7/3676">ref</a>) <ul> <li>Set up competing personas and add a "critic" to help validate or improve outputs.</li> </ul> </li> </ul> <h1 id="references">References</h1> <ul> <li><a href="https://www.youtube.com/watch?v=pwWBcsxEoLk">The Ultimate Guide to AI Prompting</a></li> <li><a href="https://www.promptingguide.ai/">Prompt Engineering Guide</a></li> </ul> <h1 id="what-should-you-do-now">What should you do now ?</h1> <p>Maybe using AI to create the prompt for you is the best thing to do :)</p> <p>For that, use the following text as an instruction for llms to write the prompt for you:</p> <p>(Think of it as a blog inside of a blog)</p> <pre data-lang="md" style="background-color:#2b303b;color:#c0c5ce;" class="language-md "><code class="language-md" data-lang="md"><span style="color:#8fa1b3;"># The Ultimate Guide to Creating Better AI Prompts </span><span> </span><span style="color:#8fa1b3;">## The Core Principle </span><span style="font-weight:bold;color:#ebcb8b;">**AI prompting is programming with words, not asking questions.**</span><span> You're not having a conversation—you're instructing a prediction engine that completes patterns based on probability. </span><span> </span><span style="color:#8fa1b3;">## The Meta-Skill: Clarity of Thought </span><span>Before you write a single prompt, </span><span style="font-weight:bold;color:#ebcb8b;">**think clearly about what you want**</span><span>. If you can't explain it clearly to yourself, you can't prompt it. All prompting problems are actually thinking problems. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## Essential Techniques </span><span> </span><span style="color:#8fa1b3;">### 1. </span><span style="font-weight:bold;color:#ebcb8b;">**Use Personas** </span><span>Tell the AI who it should be to narrow its focus and improve responses. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Example:**</span><span> "You're a senior site reliability engineer for CloudFlare. You're writing to both customers and engineers." </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Why it works:**</span><span> Gives the AI a specific perspective and expertise to draw from instead of generic knowledge. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">### 2. </span><span style="font-weight:bold;color:#ebcb8b;">**Provide Context (Most Important)** </span><span>Give the AI ALL the details it needs. Never assume it knows something—more context = fewer hallucinations. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Key principle:**</span><span> ABC = Always Be Contexting </span><span style="color:#bf616a;">- Include all relevant facts </span><span style="color:#bf616a;">- Be specific and detailed </span><span style="color:#bf616a;">- Don't hold back information </span><span style="color:#bf616a;">- Give it permission to say "I don't know" if it lacks information </span><span style="color:#bf616a;"> </span><span style="font-weight:bold;color:#ebcb8b;">**Pro tip:**</span><span> Whatever context you don't provide, the AI will fill in with guesses. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">### 3. </span><span style="font-weight:bold;color:#ebcb8b;">**Define Output Requirements** </span><span>Tell the AI exactly how you want the result to look. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Specify:** </span><span style="color:#bf616a;">- Format (bulleted list, under 200 words, etc.) </span><span style="color:#bf616a;">- Tone (professional, casual, anxious, etc.) </span><span style="color:#bf616a;">- Structure (timeline, sections, headers) </span><span style="color:#bf616a;">- Style constraints (no corporate fluff, radically transparent) </span><span style="color:#bf616a;"> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">### 4. </span><span style="font-weight:bold;color:#ebcb8b;">**Use Few-Shot Examples** </span><span>Show the AI what you want instead of just describing it. Provide 2-3 examples of the exact output style you're looking for. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Why it works:**</span><span> Removes guesswork—you're teaching by example rather than description. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## Advanced Techniques </span><span> </span><span style="color:#8fa1b3;">### </span><span style="font-weight:bold;color:#ebcb8b;">**Chain of Thought (COT)** </span><span>Ask the AI to think step-by-step before answering. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Example:**</span><span> "Before writing this email, think through it step by step..." </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Modern shortcut:**</span><span> Use "Extended Thinking" or "Reasoning" mode built into most AI platforms. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">### </span><span style="font-weight:bold;color:#ebcb8b;">**Tree of Thought (TOT)** </span><span>Have the AI explore multiple approaches simultaneously, then synthesize the best path. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Example:**</span><span> "Brainstorm three distinct approaches: [A], [B], [C]. Evaluate each, then synthesize them." </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">### </span><span style="font-weight:bold;color:#ebcb8b;">**Battle of the Bots (Adversarial Validation)** </span><span>Create competing personas that critique each other. </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Example:**</span><span> "Generate a 3-round competition with [persona 1], [persona 2], and [critic]. Have them debate and collaborate on the final output." </span><span> </span><span style="font-weight:bold;color:#ebcb8b;">**Why it works:**</span><span> AI is better at critiquing than original creation. </span><span> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## Common Mistakes to Avoid </span><span> </span><span style="color:#bf616a;">1. </span><span style="font-weight:bold;color:#ebcb8b;">**Being vague**</span><span style="color:#bf616a;"> - AI fills gaps with guesses </span><span style="color:#bf616a;">2. </span><span style="font-weight:bold;color:#ebcb8b;">**Skipping context**</span><span style="color:#bf616a;"> - Leads to hallucinations </span><span style="color:#bf616a;">3. </span><span style="font-weight:bold;color:#ebcb8b;">**Assuming knowledge**</span><span style="color:#bf616a;"> - AI doesn't remember everything </span><span style="color:#bf616a;">4. </span><span style="font-weight:bold;color:#ebcb8b;">**Not defining output**</span><span style="color:#bf616a;"> - Gets generic results </span><span style="color:#bf616a;">5. </span><span style="font-weight:bold;color:#ebcb8b;">**Messy thinking**</span><span style="color:#bf616a;"> - Creates messy prompts </span><span style="color:#bf616a;"> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## The Process </span><span> </span><span style="color:#bf616a;">1. </span><span style="font-weight:bold;color:#ebcb8b;">**Think first, prompt second** </span><span style="color:#bf616a;">2. Write down exactly what you want before prompting </span><span style="color:#bf616a;">3. Ask yourself: "Could a human do this with the information I've provided?" </span><span style="color:#bf616a;">4. If yes → the AI probably can too </span><span style="color:#bf616a;">5. If no → add more context and clarity </span><span style="color:#bf616a;"> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## Quick Wins </span><span> </span><span style="color:#bf616a;">- </span><span style="font-weight:bold;color:#ebcb8b;">**Enable web search**</span><span style="color:#bf616a;"> when you need current information </span><span style="color:#bf616a;">- </span><span style="font-weight:bold;color:#ebcb8b;">**Save successful prompts**</span><span style="color:#bf616a;"> in a library for reuse </span><span style="color:#bf616a;">- </span><span style="font-weight:bold;color:#ebcb8b;">**Use prompt enhancers**</span><span style="color:#bf616a;"> to improve your raw prompts </span><span style="color:#bf616a;">- </span><span style="font-weight:bold;color:#ebcb8b;">**Treat bad outputs as skill issues**</span><span style="color:#bf616a;"> - the problem is usually your clarity, not the AI </span><span style="color:#bf616a;"> </span><span style="background-color:#4f5b66;color:#c0c5ce;">--- </span><span> </span><span style="color:#8fa1b3;">## Remember </span><span> </span><span>The AI can only be as clear as you are. When frustrated with results, don't blame the AI—look in the mirror. It's a skill issue. Get clearer about what you want, and the AI will get clearer in its responses. </span></code></pre>