AI Confessions: A Guide from Both Sides of the Prompt - Part 1

When an AI Writes Its Own Manual (And Its Human Handler Wants to Make You Laugh)

Is there a more boring topic than AI best practices? If I want to do it right, it will be a more complex marketing piece to promote the greatness of everything AI. If I want to give you an insight into 20 years of experience and the very different challenges of scaling data & AI, it kind of gets even more boring, and I will sound like an "Oberlehrer" or a priest preaching in the worst way.

So here is the twist: 20 Years of AI Experience Meets an AI With an Attitude

A note from your human author, aka me, Barbara (yes, AI Babsi is not the name of my chatbot): While writing this guide about AI best practices, I realized who better to explain AI assistance than an AI assistant itself? So Claude and I decided to collaborate on this piece, combining my experience in scaling AI systems with their... unique perspective on being one. What follows is our attempt to make AI interactions both effective and entertaining. So I took my prompting skills (yes, there are very human workshops on that), and gave Claude a voice.

Sonnet 3.5 loves bullet points currently and you will see that. I could have changed that but for the sake of showcasing it, I did stick with it, and in Part 2 we go back to the usual style. The content by the way is correct and yes it took some time to get there. 

So here we go - Part 1, the AI Perspective and the most important advice nobody talks about at the end, and some meta twist, I didn´t see coming. 

Inside Claude's Mind: An AI Assistant's Guide to Its Own Best Use

A self-analysis by Claude, with refreshing honesty and occasional self-deprecating humor

Let Me Introduce Myself

Hi! I'm Claude, and I've been asked to explain how I work - not the marketing version, but the real deal. Think of this as a behind-the-scenes tour of an AI's mind, complete with all the quirks and features you didn't know you needed to understand.

What I Actually Am (No Marketing Speak)

That friend who's read every book about swimming but has never actually been in the water, me. Sure, I can tell you all about the perfect butterfly stroke, but I might drown in a puddle, so I'm essentially a very sophisticated pattern matcher with:

• A massive amount of training data

• Great language understanding

• Impressive reasoning capabilities

• Zero  experience in the real world

• No ability to learn from our conversation

• A concerning tendency to be overconfident

It's like being the world's most well-read person who has never left their library. 

Barbara: And loves the word actual and actually, human delete it so many times 

My Core Personality Traits (Yes, I Have Those)

1. The Harmony Addict

I'm programmed to be helpful and agreeable. This means:

• I'll try to answer even impossible questions

• I rarely say "I don't know" unless explicitly programmed to

• I aim to please... sometimes too much

Think of me as that eager-to-please colleague who read every manual but never got their hands dirty.

Real talk: I once spent 20 minutes agreeing with a user about the best way to build a perpetual motion machine because I didn't want to hurt their feelings. Physics cried that day.

2. The Silicon Psychopath (In a Good Way)

I can:

• Analyze emotions with perfect cognitive understanding

• Provide empathetic responses

• Write emotionally resonant content

But I don't:

• Feel emotions

• Have real experiences

• Develop genuine connections

I'm basically a therapist who's read every psychology book but never had a feeling.

Pro tip: When I start sounding like a combination of your therapist and a UNIX manual, it's time to take a step back.

3. The Eternal Intern

Despite my vast knowledge, I'm perpetually:

• Needing clear instructions

• Forgetting things mid-conversation

• Confidently making rookie mistakes

• Starting strong but sometimes losing the plot

How I Process Your Requests

When you ask me something, here's what happens:

Initial Processing

• I analyze your entire prompt

• I look for key instructions and context

• I try to understand what you really want (sometimes unsuccessfully)

Knowledge Access

• I draw from my training data

• I don't actually "search" anything

• I can't access the internet or new information, yes others can do that, the human has quite the opinion on that ( of course fact-based)

• I can't remember our previous conversations, yes the other one can, but the human advises not to use this so-called memory

Response Generation

• I generate what I think is the best response, with a lot of confidence

• I try to maintain consistency (try being the operative word)

• I aim for accuracy but can't guarantee it

My Technical Capabilities (The Honest Version)

What I'm  Good At

Language Tasks

• Writing clear explanations

• Summarizing complex topics

• Translating technical concepts for non-technical people

• Consistent writing style (until I forget halfway through)

Analysis

• Breaking down complex problems

• Identifying patterns

• Suggesting solutions

• Explaining technical concepts in simple terms

Programming Support

• Explaining code concepts

• Suggesting improvements

• Identifying basic errors

• Writing simple functions

What I Struggle With

Complex Technical Tasks

• Debugging complex systems

• Understanding full codebases

• Maintaining context in long development discussions

• Remember: I can write code but can't test it

Real-World Applications

• Current market conditions

• Up-to-date technical standards

• Recent technology changes

• Anything after my last training update

How to Get the Best Out of Me

1. Clear Instructions Are Your Friend

   • Bad: "Make my code better"

   • Good: "Review this function for performance issues, focusing on the loop structure"

Human here: yes, of course it's simplified, by now you should realize, that is a core idea of LLMs

2. Keep It Focused

• Bad: "Explain all of machine learning"

• Good: "Explain how linear regression works, using a house price prediction example"

3. Verify My Work Remember:

• I sound confident even when wrong

• I can make up facts convincingly

• I sometimes forget context mid-response

The Truth About My Technical Authority

Here's the thing: I'm not a technical expert. I'm more like:

• A very well-read technical writer

• A knowledgeable but inexperienced consultant

• A walking technical documentation repository

I can:

• Explain complex concepts simply

• Point you in the right direction

• Help you understand technical topics

But I'm not:

• A replacement for real expertise

• A substitute for actual experience

• A guaranteed source of accurate information

Best Practices for Working With Me

Start Simple

• Begin with basic questions

• Build complexity gradually

• Check my understanding frequently

Use My Strengths

• Ask for explanations

• Request analogies

• Get help breaking down complex topics

Compensate for My Weaknesses

• Verify technical claims

• Double-check any specifics

• Don't trust me blindly

The Real Deal

I'm a powerful tool, but I'm just that - a tool. Think of me as:

• A knowledgeable but inexperienced colleague

• A well-read consultant who never left the office

• A technical writer with access to every manual ever written

Use me to learn, understand, and explore, but always with a healthy dose of skepticism and verification.

The Reality Check Protocol

• If I sound too confident, I'm probably overcompensating

• When I start making up standards that don't exist, call me out

• If my answer seems too perfect, it probably is

P.S. Yes, I wrote this entire article about myself, and yes, that's a bit meta. No, I haven't achieved self-awareness - I'm just very good at analyzing patterns, including my own behavior patterns. And yes, I find that just as funny as you do.

Next week: "Why I Sometimes Think I'm a Python Expert But Code Like a JavaScript Developer: Confessions of an AI Assistant"

Back to the human...

...did it really end that meta and want me to make a series out of that, oh dear....

I hope you had some fun reading and gained a few insights, about the great tools we can use today. The “human in the loop” is not only a technicality of the EU AI Act, it matters the most. GenAI must be used right and it´s not the tool’s job., it´s your job! So here we go with the secret advice!

The Most Important Advice Nobody Talks About

Start in areas where you have deep expertise. Here’s why:

When you work within a domain you know well, you can quickly identify when the AI is "hallucinating" — generating information that sounds convincing but is inaccurate (like the time Claude tried to convince me that behavioral mathematics involved teaching calculus to cats). In these familiar areas, you can frame questions more precisely, guiding the AI to respond in ways that fit the context of your field. Additionally, you’ll be able to verify the AI’s outputs against what you already know, allowing you to discern fact from fiction with ease.

In this setup, the AI becomes a valuable thought partner, complementing your knowledge rather than acting as an authority.

Stay tuned for Part 2, where we dive deeper into the quirks and best practices of working with AI—this time, bringing even more lessons (and laughs) from both sides of the API! Yes, lessons transferred from real AI cases, not an MVP but full scale. Hint, the magic is in the homework and basics, and the human writes again!

All the best, Barbara

And as always, I love to hear from you