My Take on Prompt Engineering and Major AI Models

— Was ist schlecht? — Alles, was aus der Schwäche stammt.— What is evil? — Everything that comes from weakness.

My Stance on Prompt Engineering

To be frank, I believe deep specialization in prompt engineering yields diminishing returns for the average user. While it’s a fascinating field, mastering a few powerful, general principles is far more practical than getting lost in the weeds. This post is a summary of my journey: what I’ve learned, what I already knew, and my personal take on the major AI models I work with daily.

The Model Takes

Takes on models spoil faster than any other opinion on this page, so this section is dated on purpose: the current read comes first, and the September 2025 snapshot stays frozen below it — partly as a record of how fast the ground moves.

The 2026 update — July 2026

Why now? Because the ground moved again: Chinese models have surged once more, the race between GPT and Claude keeps tightening, and K3 now beats Fable outright on some fronts — at a lower price. The map needed redrawing.

Gemini: the talker

The bad news first: at coding, Gemini has started to lose the plot — still ahead of some Chinese models, though not of the best ones; more on that at the end.

The good news is what it stands on: the best search engine there is. That makes it close to unbeatable at information acquisition, with respectable synthesis on top.

But its alignment is tuned to please. It would rather look after your feelings than solve the problem outright, let alone push back on a decision — and when it lays out options, they can feel like they are there for show.

Images are decent, and handing it a video is nearly effortless, since YouTube lives in the same house. As a partner for talking through ideas, an eight out of ten. For code, no.

GPT: the right default

Disclosure before the complaints: they are personal, and for most people GPT remains the right pick — absent a hard reason against it, choose it and you will be fine. It has real advantages in compute and in what it charges, and codex has been doing honest work lately.

My problems with it are temperamental:

1. It runs on minimal assumptions — look first, act later, one thing at a time. Admirably steady, and the exact opposite of Claude, which assumes, and has opinions.
2. When we discuss research or untangle methodology, the reply runs to thousands of words as a matter of routine — organized, perhaps, but volume is volume, and tone instructions barely dent it. For a reader with ADHD, that is a real cost.
3. Its front-end work is weak and its taste sometimes falls short of Gemini’s — and since most of the world writes its front-ends with GPT, everything ends up looking the same, which makes it worse.

The other side of the ledger: its image generation is absurdly good (I draw with it for fun), and it can genuinely watch video — two things Claude cannot do at all. Claude’s image “generation” is code drawing SVG stroke by stroke, which is to say: it does not exist.

Claude: the work partner

Claude is my personal favorite, and I say so with both eyes open: tokens are brutally expensive, the limits on Fable are tight, and I actively dislike more than one of the company’s choices. But the work is the work.

At coding, Claude Code on top of a model sitting at or near SOTA across the board puts it in a league of its own. I know the criticisms in circulation, and I share some of them — but I use it to get work done, not to worship a company.

And for the part of my day that matters most — talking through research directions, organizing and integrating information, arguing a point all the way down, learning something hard — it is the strongest I have used. It has carried real weight in how I sort out my research decisions. Its deep research is genuinely decent; at raw search, Google still wins.

Tools, not idols

One last point, probably the only one in this section that will not expire: do not turn models into fandoms. This is not idol-chasing. They are tools — for me, tools and working partners at once, which is exactly why a model that argues back with a defensible take of its own is worth more than one that only agrees.

No company worship either: the moment a tool stops earning its keep, replace it, same day. Concretely — Kimi K3 is already an Opus-class model, and at web dev it beats even Fable 5 by a clear margin. The day it talks research at Claude’s level for a third of the price, I switch.

Frozen: the September 2025 snapshot

It’s clear that not all models are created equal. Different models excel at different tasks, and understanding their strengths is key to using them effectively. Here are my personal observations.

Summary

Claude: The Coder

In my experience, Claude is a top-tier model for coding tasks. Its ability to understand and generate code is exceptional. A standout feature is its remarkably low hallucination rate — it rarely invents facts, which builds a high degree of trust. However, this strength in coding is sometimes a double-edged sword. It can be prone to over-engineering solutions by adding unrequested features without mentioning them. For instance, when asked for a script to simply transmit raw BVP sensor data, it preemptively added logic to calculate and send heart rate as well. While its writing capabilities are also strong, its performance in brainstorming and decision-making seems weaker, often favoring “safe” or middle-of-the-road answers.

GPT: The Brainstormer

I find GPT to be a strong contender for brainstorming and ideation, rivaling others in its ability to generate creative options and explore different angles of a topic. However, it seems to struggle with vetting the credibility of online sources, often accepting what it finds at face value. Similar to Gemini, it can be prone to hallucinations without self-awareness and also shares Claude’s habit of trying to anticipate user needs, sometimes adding unnecessary complexity. While its coding abilities are solid, I feel they don’t quite reach Claude’s level. It’s a capable all-rounder but hasn’t become my go-to for specialized tasks.

Gemini: The Synthesizer

From my perspective, Gemini excels at search-related queries and synthesizing information, with brainstorming capabilities that are on par with GPT’s. In contrast to other models, Gemini seems to have a better grasp of what a user explicitly asks for, sticking closer to the prompt’s requirements.

Regarding its coding abilities, while the raw output may not always match the complexity of Claude’s, its greatest strength is its predictability and reliability. It follows instructions with high fidelity, making it a valuable partner when you need the output to strictly match the requirements. In my long-term coding work with Gemini 2.5 Pro, its most impressive trait is its long-context memory. It demonstrates a remarkable ability to recall details from earlier in a complex, extended conversation — a significant advantage for iterative development.

The main drawback, however, is its tendency to hallucinate. This can manifest in subtle ways, but sometimes it’s stark. For instance, when asked to find research papers, it might invent one with a plausible title, authors, and a fake URL that leads nowhere.

Practical Techniques That Actually Work

Even with a minimalist philosophy, I’ve picked up a couple of effective techniques that have notably improved my AI interactions.

1. Use HTML-like Tags for Structure

I used to rely on code blocks to structure my prompts. However, I’ve found that using HTML-like tags is a more robust method to delineate different parts of a prompt. It helps the AI clearly distinguish between instructions, references, and examples.

For example:

<instructions>
Please analyze the following user feedback.
</instructions>
<feedback_text>
The user reported that the application crashes on startup.
</feedback_text>

2. AI’s Persona

Assigning a persona to the AI is a technique I never made a habit of, but it can be surprisingly effective. It’s more than just a gimmick; it sets a specific context, tone, and knowledge level for the AI to emulate.

For instance:

“Assume you are a seasoned professor of AI research. Please guide me through the core concepts of reinforcement learning.”

Core Principles That Always Matter

Some aspects of interacting with AI have been clear from the start and remain fundamental to getting quality results.

1. Maximize Signal, Minimize Noise

Context is king, but relevant context is the kingdom. Provide the model with what it needs to succeed, not just everything you have. State the goal, constraints, and key assumptions. When asking for code, include the function’s purpose, its API, and where it fits in the system.

Conversely, “polluting” the context with irrelevant information, like a blurry photo of a screen for a debugging session, is counterproductive.

2. Deconstruct Large Problems

Expecting an AI to handle a vague, large-scale request in one go is a recipe for mediocrity. Breaking down a complex task into smaller, sequential steps is a far superior strategy.

For instance, instead of asking to “create a presentation about Topic X,” a better approach is to break it down:

1. Outline the key sections for a presentation on Topic X.
2. Draft the content for each slide.
3. Suggest a visual theme and layout.
4. Generate speaker notes.

If you don’t know the steps, you can ask the AI to propose a plan.
Still, applying your own domain knowledge to deconstruct the task will always produce a superior outcome.

My Practical Checklist

Here’s a distilled checklist of the tactics I find myself using daily.

1. Structure First: Use tags (<goal>, <context>, <constraints>) to clearly delineate the components of your prompt.
2. Persona on Purpose: Only assign a persona when it meaningfully adds constraints or provides a useful frame of reference (e.g., “act as a skeptical code reviewer”).
3. Plan Before Execution: For complex tasks, ask the AI for a concise plan or outline first. Agreeing on the structure prevents wasted effort.
4. Evidence-First for Facts:
   • Insist on citations with URLs before the AI makes a claim.
   • For research tasks, ask for an “evidence table” (source, claim, URL) before the final synthesis.
5. Specify the Output: Clearly define the desired format, length, style, and any acceptance criteria the output must meet.
6. Iterate Smartly: Work in small loops. Freeze what’s working and change only one variable at a time to isolate what improves the result.

Debatable “Best Practices”

There are a few commonly cited prompting techniques that I’m not yet convinced of.

• Focus on “Dos” over “Don’ts”: The idea that you should always frame requests positively.
• Use a respectful tone: The theory that being polite to the AI yields better results.

My experience hasn’t shown these to have a consistent, measurable impact, but I remain open to the possibility.

Conclusion

In the end, effective interaction with AI is less about secret handshakes and more about clarity, context, and choosing the right tool for the job. While the models evolve, the core principles of clear communication and structured thinking remain timeless.