The Secret Life of AI: Why Your Best Prompts Fail

How to prompt, think, and get results from any AI tool

#WorkingWithAI #Prompting #AIConfidence

Margaret is a senior software engineer. Timothy is her junior colleague. They work in a grand Victorian library in London — and in every episode, they'll show you exactly how to get what you want from AI.

Episode 6

Timothy set his notebook on the table with the quiet precision of someone who had prepared carefully and was still confused about what had gone wrong.

"I did everything right," he said.

Margaret looked up. She recognized that particular frustration — not the hot anger of the rage quit, but the cooler, more bewildering feeling of having followed the instructions and still arrived nowhere useful.

"Tell me," she said.

"I spent twenty minutes on the prompt." He opened the notebook. "Specific. Clear. Structured. The kind of prompt you'd hold up as an example." He paused. "The output was useless. Generic. Like it hadn't heard a word I said."

Margaret nodded slowly. "What were you asking it to help you with?"

"A proposal. For a client I've been working with for two years."

She was quiet for a moment. Then: "And how much of those two years did you put in the prompt?"

Timothy looked at her. "What do you mean?"

"The client," she said. "Their industry. Their particular concerns. The history between you. The tone they respond to. The thing they almost said no to last time." She folded her hands. "How much of that was in your prompt?"

He said nothing.

"That," Margaret said, "is context."

The AI Doesn't Know What You Know

She stood and walked to the shelves, running a finger along a row of spines without pulling anything out. It was a habit Timothy had come to recognize — she thought better when she was moving.

"Here is the thing most people never quite grasp," she said. "You walk into a conversation with an AI carrying everything you know. Two years of client history. Industry nuance. Unspoken preferences. Lessons from previous failures. All of it present in your mind, shaping every word you type."

She turned to face him.

"The AI walks in with none of it."

Timothy leaned back. "So the prompt wasn't the problem."

"The prompt was fine. The briefing was missing." She returned to her chair. "There is a difference between asking a good question and giving someone enough to actually answer it well. You asked a good question. You assumed the answer lived somewhere in the question itself. It didn't."

"Because the AI doesn't know what I know."

"It knows what you tell it," Margaret said. "Nothing more. Nothing less. Every assumption you carry silently into the conversation — every piece of background you consider obvious — is invisible to the tool. It cannot read the two years. It can only read the prompt."

Timothy was quiet for a moment, turning this over.

"So the prompt is only half the job."

"The smaller half," Margaret said. "On a good day."

Give the AI Enough to Work With

She picked up her pen and wrote two lines on a piece of paper, then handed it to Timothy. He read it slowly.

What do you want?
What does the AI need to know to give it to you?

"Most people answer the first and skip the second entirely," she said. "They treat the AI like a search engine — type the query, receive the answer. But a search engine retrieves. The AI reasons. And reasoning requires material to work with."

Timothy studied the two lines. "So before I write the prompt I should be writing the context."

"Before you write the prompt you should be asking yourself: what does this tool not know that it needs to know?" She tapped the paper. "The client's industry. The audience for the proposal. The tone that has worked before. The constraints you're working within. The thing you're trying to avoid. Give it the briefing, not just the question."

Timothy nodded slowly. "So instead of 'write a proposal for a cloud migration' it's 'write a proposal for a client who is terrified of downtime, loves bullet points, and rejected our last bid because it felt too corporate.'"

"Now you're giving it something to chew on," Margaret said.

"And if I had done that —"

"You would have gotten something useful," Margaret said simply. "Not because the prompt was better. Because the AI finally had enough to work with."

Giving the AI a Proper Briefing

Timothy turned to a fresh page in his notebook.

"How much context is enough?"

Margaret considered this. "Enough that a capable stranger — someone intelligent, willing to help, but knowing nothing about your situation — could do something useful with it." She paused. "That is essentially what you are working with every time. A capable stranger. Treat it accordingly."

"So if I wouldn't expect a new colleague to produce something brilliant on their first day without a proper briefing —"

"Then you should not expect the AI to do it either." She smiled. "The tool is not failing you when it gives you something generic in response to a context-free prompt. It is doing exactly what it can with what it has. Which is very little, if you gave it very little."

Timothy wrote that down. Then looked up.

"The frustrating part is that it sounds so confident even when it has nothing to work with."

"Yes," Margaret said. "That is a separate problem for a future conversation." She picked up her book again. "For now — next time you sit down with the tool, try writing the context before you write the prompt. Everything the AI doesn't know but needs to. Then the prompt."

"And if it still goes wrong?"

She looked at him over the top of her book.

"Then at least you'll know it wasn't the context."

The Secret Life of AI is a series about the human side of working with artificial intelligence — the skills, habits, and mindset that determine whether the tool works for you or against you. Because the missing piece was never the AI.

Next episode: The AI sounds certain. That doesn't mean it's right.

Aaron Rose is a software engineer and technology writer at tech-reader.blog.

Catch up on the latest explainer videos, podcasts, and industry discussions below.

Search This Blog

Tech-Reader.blog