Your AI Isn't the Problem. Your First Prompt Is.

Most people treat ChatGPT like a calculator and get frustrated when it doesn't act like one. One founder's five-step fix, tested over two years of building custom AI assistants.

Welcome Back,

We've all done it, built a custom GPT, fed it documents, and shouted out in frustration when it failed real-world tasks, hallucinating, ignoring a key rule, or giving a generic, useless answer, according to entrepreneur and Kalicube founder Jason Barnard, writing for Entrepreneur. His diagnosis is blunt, if that's happening to you, the technology is up to the task, your approach isn't.

Most entrepreneurs are treating ChatGPT, Gemini, and Claude as software to be configured, when they're closer to a bonkers mix of super-knowledgeable interns that need to be guided. Writing a static set of rules and expecting flawless performance is, in Barnard's words, a huge mistake. Over the last two years, he's built dozens of specialized AI assistants, moving from frustratingly unhelpful tools that drown him in useless text to genuinely powerful ones that save real time.

Today we look at the two traps most people fall into when building an AI assistant, the specific five-part job brief that replaces a vague prompt, why a flawed first draft is actually useful information, and the one step Barnard says almost everyone skips, the step he calls the single biggest source of AI frustration.

⚠️ The Two Traps Everyone Falls Into

When building a custom AI assistant, most entrepreneurs fall into one of two traps, according to Barnard. The first is the one-line prayer, a vague prompt that expects the machine to read your mind. The second is the mega-manual, a single, massive prompt that tries to account for every possible scenario upfront.

Both create an inflexible system that breaks under pressure, forcing a frustrating cycle of rewrites. The reason, according to Barnard, is that both methods treat the AI like a calculator, not a collaborator. The AI has no context for why the rules exist, it's just executing a script.

That reframing is genuinely useful on its own, independent of the specific five-step method that follows. Most frustration with AI assistants comes from expecting a single, perfect instruction set to work immediately, the same unrealistic expectation nobody would place on an actual new hire's first day.

Treating an AI assistant like a new hire, one that needs coaching, not a finished configuration, is the core mindset shift.

📋 Step 1: Ditch the Prompt, Write a Job Brief

The solution, Barnard argues, is to treat your AI assistant like a new hire. Your first prompt should be a clear job brief, written with a plain, unconfigured version of Gemini or GPT, defining the core components in a single, focused conversation. He lists five components every brief needs.

Barnard adds a sixth element worth calling out separately, clarifications, an instruction empowering the assistant to ask for help. Something as simple as "if any critical information is missing to meet the goal, ask me targeted questions before you proceed" gives the AI permission to flag a gap instead of quietly guessing and producing a confidently wrong answer.

✍️ Step 2: Assign a Real Task, Expect a Flawed First Draft

Once the assistant has its brief, give it its first real-world assignment, a real input it's never seen before, and see what it produces. The expectation matters here as much as the task itself, expect it to get things wrong. Just like a junior employee, its first attempt will reveal the gaps in its understanding.

That reframing of a bad first draft as useful data, not wasted effort, is a genuinely healthy way to approach this. A flawed output isn't a sign the process failed, it's the diagnostic step that tells you exactly what needs correcting next, the same way a new employee's early mistakes reveal exactly what training they still need.

A flawed first draft is diagnostic information, not a failure, revealing exactly which gaps the AI's instructions still need to close.

🗣️ Step 3: Feedback, Not a New Set of Rules

This is where the real coaching begins, and Barnard is specific about the structure that works. Present the critique as, first, a reminder of the input you provided, second, the output the AI generated, and third, your actual critique, stated specifically. His example, "The tone was excellent, but you failed to mention the key statistic about Q3 revenue, which was a critical part of the input. That statistic needs to be in the first paragraph."

That last instruction is the subtle, important part, rather than the user rewriting the rules themselves, the AI is asked to update its own instructions based on the specific feedback given. Barnard recommends staying in a single conversation throughout this phase, so the model is building on existing context, adding new rules, refining old ones, and, crucially, learning why the previous version fell short, not just what changed.

🔍 Step 4: The Step Almost Everyone Skips

Here's the step Barnard calls the most important, and the one he says most people miss entirely. When you ask an AI to "tweak" its instructions, it can sometimes remove or alter foundational rules in the process of incorporating new feedback, a pattern he calls "instructional drift."

Barnard's own recovery prompt is worth stealing directly, "You have removed the rule about X. That is still a core requirement. Please add it back in, and then tell me what other essential instructions you may have omitted in your update." That second half of the sentence does real work, it turns the correction into an audit, rather than a one-off patch.

Comparing updated instructions against the previous version catches rules that quietly disappeared during a "tweak."

🧠 AI Spotlight Analysis

The fifth and final step is simply to repeat the cycle, assign, review, critique, and audit, until the AI's understanding becomes more nuanced and its performance more reliable. But Barnard adds one last, genuinely counterintuitive piece of advice at the end, once the instructions are finally right, start a brand new assistant with them, rather than continuing to use the one you've been training.

"The assistant you have been working with will still contain all the inaccuracies, mistakes and back-and-forth from steps 1 to 4. That pollutes the algorithm," Barnard writes, calling this "historical assistant pollution," and describing it as the single thing that makes custom GPTs so frustrating for most people. That's a genuinely useful, easy-to-miss detail, the messy training process itself leaves residue behind, and the clean, refined instructions deserve a clean slate to actually perform on.

That's a genuinely useful reframe for anyone who's given up on a "frustrating" AI assistant after one bad session. The method here isn't exotic, it's structured iteration, treating the model less like software you configure once and more like a new hire you coach through repeated, specific feedback. The upfront patience required is real, but so is the payoff Barnard describes once "historical pollution" is finally cleared out of the picture.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.