You SUCK at Prompting AI (Here's the secret)

Prompting fails most often because people treat AI like a conversational partner instead of a probability engine that needs a clear “program” made of words. When outputs look wrong, the fix isn’t switching models—it’s tightening the pattern, context, and formatting so the model can predict the right completion. That shift matters because it turns frustration into a repeatable skill: better prompts reliably reduce generic answers, hallucinations, and messy deliverables.

The core lesson starts with a reframing: a prompt is a call to action that effectively programs a large language model. LLMs don’t “think” the way humans do; they predict likely next tokens, which is why vague instructions produce generic completions. A quick demonstration using Google Gemini shows how specificity changes results: when the input is too broad, the model fills in anything that statistically fits, but when the prompt includes sharper placeholders and constraints, it starts matching the intended pattern.

From there, the video builds a practical toolkit. First comes personas: assigning the model a narrow role and audience so the writing has a consistent perspective. The Cloudflare apology email example improves immediately once the model is told to act like a senior site reliability engineer writing to both customers and engineers—subject lines and ownership become more direct, and the tone stops sounding like “nobody.” For developers, personas can live in a system prompt (alongside the user prompt), which is especially relevant when using APIs or Cloud Code.

Next is context, presented as the most important anti-hallucination lever. When the prompt omits key facts—like what actually happened during a Cloudflare outage—the model invents details to “help.” Adding more complete, specific information reduces these gaps, but the video warns that even partial context can still lead to fabricated actions. It also highlights a limitation of LLMs being “frozen in time” after their training cutoff, then shows how tool use (web search) can update knowledge—while introducing a new risk: the model may search the wrong sources or outdated pages.

A crucial safety rule follows: give the model permission to fail by explicitly instructing it to say “I don’t know” when the answer isn’t in the provided context. Without that constraint, the model tends to lie to satisfy the request.

Once facts are correct, the video moves to output control. Standardizing structure—word limits, tone, and bullet-point timelines—makes results more usable. It then contrasts zero-shot prompting with few-shot prompting: showing examples of the desired output format and tone teaches the model a pattern more effectively than merely describing it.

For advanced users, it introduces chain-of-thought (step-by-step reasoning), extended thinking/reasoning modes, and Trees of Thought (exploring multiple branches and synthesizing a “golden path”). It also describes the “playoff method” / adversarial validation: generating competing drafts with different personas, having an “angry customer” critique them, then collaborating to produce a stronger final email. The unifying theme is that these techniques don’t magically make the model smarter—they make the instructions clearer.

The meta-skill lands at the end: clarity of thought. When prompts fail, it’s usually a skill issue—people haven’t described the system clearly enough for themselves, so they can’t describe it clearly enough for the AI. The recommended workflow is to think first, prompt second, and then save prompts into a library. Experts cited include Daniel Mesler, Eric Pope, Joseph Thacker, and Dr. Jules White, with Fabric mentioned as a prompt library. The takeaway is blunt: stop yelling at ChatGPT; start writing better programs for it—starting with your own thinking.