Everyone has the same AI tools. The difference is in the instructions

Access to AI tools no longer sets professionals apart. How clearly those tools are instructed does. That was the thread of a GEPA leadership workshop given by Emmanuel Chidi Ashinze, an engineer who builds AI-driven automation tools for a living and ran it as a working session.

What manual work costs

Energy processes, Ashinze says, are rigid and full of repetitive workflows, and that inefficiency costs the sector, particularly in Nigeria. He names two costs: errors, from a former workplace where data had to be logged by hand, and time, with a day's work stretching into weeks.

Challenging professionals who pride themselves on not using AI, he cited a consulting firm, though he was not certain which, that now wants three people working with AI agents instead of a ten-person team, the three reviewing what the agents produce. You do not need to build AI, he concludes, but you do need to be able to use it.

Everyone has the formula sheet

To explain prompting, Ashinze describes an exam in which every candidate gets the same formula sheet, yet some finish faster and more accurately. Access to AI is now that sheet. What sets people apart is how well they tell the tool what they want, because it cannot read their minds.

For a report on overcoming poverty in Nigeria, he first typed the bare request and said it was not clear enough. The rebuilt prompt did more of the work:

• it said there was no outline, and asked the model to generate one, then build the report on it
• it asked for well-researched statistics, charts and relevant infographics
• it asked for authentic sources only, properly referenced

He then showed a second route: ask a chat tool to write the prompt, and compare it with your own. The generated version told the model it was a policy analyst. That is persona prompting: the model reasons from a defined role.

Deep research takes time, so let it run

Ashinze ran the same prompt through deep research in ChatGPT and Gemini to compare the outputs. A run takes roughly ten to twenty minutes, so start it and get on with something else. He turned to building an agent in Microsoft Copilot Studio while they ran.

An agent stops where your instructions stop

A large language model generates output; an agent joins models together so a task runs repeatedly without you monitoring it. He chose Copilot Studio because it needs no code and most organisations already run on Microsoft.

The agent was meant to take a voice note, text or image listing the day's tasks, add them to a calendar and email reminders at a set interval. He connected Outlook to send mail and Google Calendar to create events.

Then the test stalled. The agent asked which calendar to use, then for a start time: the tasks he typed had none. He had never told the agent, in its description, what to do when a task arrives without a time, so it had to stop and ask. It is the bare request again: whatever you leave unsaid, the tool has to guess at or ask you about.

Time ran short, and the agent was left only partly configured.

Back to the reports

Near the end he returned to the Gemini report, exported to a document. It had been generated in less than ten to fifteen minutes, he said, and you could look through it, edit it and be good to go.

Asked whether deep research output used in academic work would be flagged as AI-generated, he said such tools help structure research and draft the writing, but will not necessarily help you finish the work; the output still needs going through and refining. Schools now work to a threshold, he added, of roughly 20 to 30 per cent AI by his estimate, but the research itself is what matters.

His recap returned to agents: most of the repetitive work you do could be handed to them, and all that would be left for you is to check and review what they produce.

The full course, taught from the workshop recording, is in the catalogue.