axcdev

AI, But How?

2025-08-28

  • Blog
  • Artificial Intelligence (AI)
  • Large Language Models (LLMs)
  • Model Context Protocol (MCP)

With the release of every new LLM, we’re promised the world. “PhD-level agents” as Sam Altman put it. But where are they? And more importantly, how do you get one yourself?

Unfortunately, it’s very likely that you are the limiting factor in how powerful AI feels to you.

Anyone who watched the GPT-5 release stream (but not too closely) knows how amazing these model canbe. So, naturally, you open chatgpt.com, type in a request, and wait for magic to happen.

And then… disaster.

It gives you a result, sure, but not the one you were hoping for. You frown, close the tab, and conclude that AI is overhyped. But is it?

AI Is Only as Good as How You Use It

AI isn’t Google.

If you were online before the first LLMs dropped, you probably remember hacking Google search like a pro:

  • Quoting phrases “like this”
  • Using a minus sign -keyword to exclude results
  • Adding site:example.com to laser-target your search
  • Filtering for PDFs or docs with filetype:pdf

You learned the quirks of search and became fluent in shaping queries that gave you the best results. But now, we bring those same short, vague habits to AI and feel frustrated when the answers don’t measure up.

LLMs are different. Not quite a search engine, but rather a conversationalist that guesses what you want to hear.

In the Google era, you did most of the heavy lifting. You searched, found the relevant pieces, and assembled them into something useful. Your results depended on your ability to understand the problem and synthesize a solution.

LLMs flip that process. The model is now doing the synthesis for you. But if it doesn’t know your goals, constraints, or background, it’s forced to guess. And it rarely guesses right.

Context, Context, Context

Context Window: The maximum amount of text (tokens) a model can process and “remember” at once.

Context is the fuel for AI. Without it, even the most advanced models feel underwhelming.

If you want exceptional output, you need to start thinking differently. Instead of asking, “What’s the shortest way to phrase my question?” ask:

“What context does the model need to do this well?”

Background information, example outputs, step-by-step instructions, and clear goals make all the difference. The more relevant context you provide, the better the model performs.

There is a caveat though, not all context is created equal. Like this video explains, LLMs struggle with ambiguity and distractions.

When you paste in a long document with your query, there’s a good chance that the model will get lost in the noise. I’ve experienced this quite often, where the model ignored my query completely and gave me something based just on the context I fed it.

Context needs to be clear and concise. Just like how it’s hard for you to find the point of a long rambly email, it’s hard for LLMs to stay on track when they’re given a lot of extraneous information.

Prompt Engineering: Your First Superpower

Prompt engineering is just communication skills rebranded for LLMs.

Imagine a stranger coming up to you and asking you: “What should I wear tomorrow?” You check the weather forecast, see that it’s going to rain, and suggest a raincoat. They then proceed to tell you that that’s not a suitable outfit for a wedding and leave. How were you supposed to know? You were treated like an LLM. No context, no examples, minimal instructions.

Now how do you take your existing communication skills and apply them to LLMs?

1. Give It a Role

LLMs don’t have a background like a person does. Telling them what role they fill is the first step:

You are a senior systems engineer specializing in LLM optimization.
Your job: review my code and improve memory efficiency.

2. Set Constraints

Tell the model what kind of output you want. Are you trying to write a report? Draft an email to a supervisor? How many words? What language? Style? The clearer your instructions, the closer the model will get to your goal.

Summarize this report in 5 bullet points.
Maximum 50 words each. No filler.

3. Work Iteratively

Working with someone that you have a relationship with is easier than working with a stranger. You know each other, have background knowledge of their situation, and can provide feedback tailored to your needs.

LLMs don’t know you. They don’t know your background. They don’t know your goals. They just need to know what to do. It’s just like working with a new friend.

Tell them about the problem that you are facing. Give them examples of what you want to achieve. Once it gives you a result, give them feedback and refine the result. Each step helps the model understand your goals and constraints and improve its output.

Most current providers help you out with this process. Cross-chat memory and integrations are all tools to help the model know more about you such that you don’t have to constantly rebuild the same context.

4. Show, Don’t Just Tell

Examples help the model understand your expectations. This is especially important for style.

Here's an example of the writing style I want.
Use this tone in your response.

Prompting well isn’t about clever hacks. It’s about providing clarity, structure, and purpose.

Going Beyond “Just Chatting”

LLMs are mostly used as chatbots: type a question, get an answer. While this is certainly what they do best, most AI organizations are working towards a more powerful future. You’ve probably heard of the term “AI agents” being passed around. But what is it?

AI Agent: An autonomous system (usually using LLMs) that can take action on its own.

Here’s how to get these (not so) magical agents for yourself.

Connect It to Your World

Imagine being able to:

  • Summarize your entire inbox in seconds.
  • Generate meeting notes automatically after every call.
  • Turn all your documents, emails, and chats into a single searchable AI-powered workspace.

Most providers already have this capability in some form. ChatGPT has connectors, Claude Desktop has Model Context Protocol (MCP), and Perplexity has the Comet browser. All of these services allow LLMs to connect and work with your personal data and tools.

But How Good are They?

Short answer: not very good.

Back to Sam Altman, he has recently expressed that he believes that investors are “overexcited about AI” and while the technology is transformative, it’s not what most believe (or hope). This is backed up by an MIT study, The GenAI Divide: State of AI in Business 2025, that showed that out of 300 public AI initiatives covering some $30–40 billion dollars in spending, only 5% showed measurable financial return for companies.

While benchmark scores are growing with each release, concerns have been raised about the quality of the results. There are concerns that the leading companies are gaming the system, optimizing their models to get a higher score.

As this research shows, most of these AI agents simply don’t perform in the real world.

Through extensive evaluation of leading LLMs, we find that even SOTA models such as GPT-5 (43.72%), Grok-4 (33.33%) and Claude-4.0-Sonnet (29.44%) exhibit significant performance limitations. In addition, our benchmark poses a significant long-context challenge for LLM agents, as the number of input tokens increases rapidly with the number of interaction steps. Moreover, it introduces an unknown-tools challenge, as LLM agents often lack familiarity with the precise usage of the MCP servers.

Model Context Protocol (MCP), what is it anyway?

MCP is an emerging standard that allows LLMs to use tools: calendars, emails, project boards, and more. Instead of being a disconnected Q&A machine, the model becomes an assistant that can act on your behalf.

For example:

“Summarize the last 10 threads from my Gmail labeled Clients and draft polite replies to the three most urgent ones.”

Without MCP, that’s impossible. With MCP, it’s a single query.

So, AI… But How?

Here’s the cheat sheet:

  • Be specific. Vagueness gets vague results.
  • Treat AI as a partner, not a magic trick.
  • Connect it to your personal tools using MCP to unlock its full potential.

AI isn’t replacing you. But the people who learn how to use it effectively?
They might.