The opinions of contributing entrepreneurs are their own. </p><div>
Key Takeaways
- To build an AI assistant that really helps you run your business, ditch the “simplistic brief” and start with a clear job description.
- Assign a real task and expect the first version to be flawed. The first attempt will reveal the gaps in understanding.
- Provide specific, actionable feedback – no new rules. Then review the changes to avoid ‘instruction drift’.
- Repeat the cycle to build a true expert. With each cycle, the AI’s understanding becomes more nuanced and its performance more reliable.
We’ve all done it. Built a custom GPT, fed documents into it, and screamed in frustration when it didn’t work in the real world. It hallucinates, ignores a key rule, or gives a generic, useless answer.
If that’s the case for you, think about it: “The technology is up to the task. My approach is not.”
Most entrepreneurs treat ChatGPT, Gemini and Claude as software to be configured… but it’s a crazy mix of super-informed interns to mentor. Writing a static set of rules and expecting flawless performance is a HUGE mistake.
Over the past two years I’ve built dozens of specialized ones AI assistantsfrom frustratingly useless tools that drown me in streams of useless text to truly powerful tools that help me move forward faster and more effectively than I ever thought possible.
The secret to building AI assistants that are truly helpful (and a HUGE time saver) is to let go of the one-shot prompt and adopt an iterative coaching approach. You need to treat your AI assistant like a brilliant, incredibly fast and hugely knowledgeable new employee who has no real work experience.
This method transforms a gimmicky chatbot into an assistant that actually helps you run your business.
Here is the exact five-step process I use.
Related: This Revolutionary Tool Will Streamline Your Sales and Workload – Here’s How to Start Using It Today.
Step 1: Ditch the “simplistic prompt.” Start with a clear job description.
When building a custom AI assistant, most entrepreneurs fall into one of two pitfalls. 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 take into account every possible scenario.
Both create an inflexible system that breaks under pressure, leaving you in a frustrating cycle of rewrites.
These methods fail because they treat the AI as a calculator and not as an employee. The AI has no context for it Why the rules exist; it just runs a script. True expertise, both human and artificial, comes from understanding the principles behind the rules. You build that insight through practice, feedback and refinement – not by handing over an encyclopedia on the first day.
The solution is to treat your AI assistant like a new hire. Your first assignment should be a clear assignment letter. Write this first briefing using a ‘vanilla’ version of Gemini or GPT and define the core components in a single, focused conversation.
Your briefing should include these five core components:
- The role: Who is the AI? (for example: “You are a junior copywriter, specialized in financial services.”)
The goal: Be clear about the most important goal. What is the one thing this output is supposed to achieve? (for example: “The goal is to write an article that convinces potential investors of our growth in the third quarter.”)
Success criteria: How will you assess the work? List two to three measurable controls that the final output must meet. This is the AI quality checklist. (e.g. “Success Criteria: 1. The article must mention the 34% increase in sales. 2. The tone must be formal and confident. 3. The last paragraph must include a call to action.”)
Limits: Determine the rules of the game. Specify the scope (for example: “Focus only on the financial results, do not mention personnel changes”) and any permitted or prohibited sources of information.
Output format: Be specific about the final product. Should it be bullet points, a formal report, or a certain number of paragraphs? Always include a number of words.
Clarifications: Give your assistant the opportunity to ask for help. Include a simple instruction, such as: “If there is critical information missing to achieve the goal, ask me specific questions before proceeding.”
Framing the task with this level of clarity gives the AI a framework for success, as well as the ‘obvious’ instructions. From the very first interaction, you are setting yourself up for profit.
Related: How to Train AI to Actually Understand Your Business
Step 2: Assign a real task and expect the first draft to be flawed
Give your newly informed assistant his first real assignment. Give it a real-world input it’s never seen before and see what it produces.
Expect things to go wrong. As with a junior employee, his first attempt will reveal the gaps in his understanding. Your job is to help improve it by identifying these gaps and suggesting solutions.
Step 3: Provide specific, actionable feedback – not a new set of rules
This is where the real coaching begins. Present your criticism in a structured way:
- Remind him of the input you provided.
Add the output it generated.
Give your criticism. Be specific. For example: “The tone was excellent, but you neglected to mention the key metric about third quarter revenue, which was a critical part of the input. That metric should be in the first paragraph.”
Ask him to rewrite his own instructions to include this feedback.
During this feedback loop phase, stick to one conversation: you’ll build on the AI’s existing context. You add new rules, refine existing rules and (vitally) teach them Why the old ones weren’t good enough. It leverages current conversation and knowledge and provides you with the most effective improvements to the fundamental instructions.
Step 4: Review the changes to avoid “instruction drift”.
This is the step most people miss – and it’s the most important. When you ask an AI to ‘adjust’ its instructions, it can sometimes delete or change fundamental rules in the process of incorporating your new feedback. I call this “instructional drive.”
Don’t rely on the assistant to self-regulate. Do a quick audit. Compare the newly generated instructions with the previous version. If you discover a missing line, explicitly correct the AI.
I use a simple prompt for this: “You removed the line about
This forces the AI to control its own work and reinforces the importance of its fundamental rules. It’s the equivalent of reminding a new employee, “Great, you’ve incorporated the new feedback, but don’t forget the basics we talked about this morning.”
Related: From co-pilot to colleague: where the AI Assistant journey is heading
Step 5: Repeat the cycle to build a true expert
This iterative process of assigning, assessing, critiquing, and auditing is how you build a truly useful assistant. With each cycle, the AI’s understanding becomes more nuanced and its performance more reliable.
Important point: Once you have the perfect instructions, create a new assistant. The assistant you worked with still contains all the inaccuracies, errors, and back and forth from steps 1 through 4. That clutters the algorithm. So start a fresh new assistant without all that ‘baggage’. You’ll thank me for this advice – in my experience “historical assistant pollution” is THE only thing that makes custom GPT/AI assistants so frustrating.
At Kalicube, we use this very methodical training to build highly effective AI assistants (ChatGPT, Gemini, etc.) for our customers that actually improve workflow and save time.
This process works. It takes a lot of patience upfront, but the reward is HUGE, so my advice is to persevere; the assistant you get at the end of this process will be the biggest productivity gain you’ve ever experienced.