Blog

I recently started using Claude Code for my development work, and I have to admit, it is without a doubt the most powerful AI coding assistant I have ever tried. It sets an entirely new precedent for what agentic AI can achieve. The sheer competence of it caught me off-guard. You give it a prompt whilst connecting it to your project, and it effortlessly navigates through all the files you give it access to and spits out its logic in seconds.

Witnessing the steadfastness of completing tasks that would normally take hours to be resolved in a matter of minutes leaves you constantly wanting more, pulling you into a different mindset where you realise that truly anything is now possible.

This exact power has triggered a profound doubt. When AI can handle the intricate architecture and the tedious implementation with little to no thought required on my part, the following question emerges: What use is there for me as a developer?

It reminds me of a scene from the very first episode of the popular 80s classic, Knight Rider. During one of his initial exchanges with Devon Miles after test-driving KITT, Michael shares his unease over the car's terrifying level of autonomy:

Michael: Oh, great. You mean it can decide to take off and go for gas, or a car wash. Just like that? Well, that would be terrific if I happened to be working under it.
Devon: It wouldn't do anything to harm you, I assure you.

Like Michael, we are sitting behind the wheel of a machine that seems fully capable of driving itself. AI might not "harm" us in the physical sense, but if we aren't careful, letting it take over entirely can quietly crush the craft we have worked on for years.

For the past couple of years, the developer world has been swept up in the era of vibecoding — the practice of letting Large Language Models (LLMs) do the heavy lifting while we sit back and play the role of high-level orchestrators. But as tools like Claude Code, Gemini and Copilot continue to push the boundaries of autonomy, I've come to a quiet realisation: AI is actively eroding our ability to code for ourselves.

Getting High On AI

The constant influx of new capabilities is not just exhausting to keep up with; it slowly chips away at the parts of the job we genuinely love. When the machine does all the thinking, we stop being developers and become prompt-aholics. Writing a story to get the perfect output becomes our only real skill, leaving the underlying code as a black box we no longer care to understand.

I am concerned at the potential loss of the muscle memory we have acquired over the years where we have been able to patiently read through class libraries, trace logic through multiple files, and meticulously debug a stubborn issue. Like a drug-addict on crack cocaine waiting for the next hit, we are trading our skill of problem-solving for a quick hit of instantly generated code.

Delusions of Grandeur

AI coding agents are remarkably good at writing units of code that look perfect in isolation. If you need a specific algorithm or a standard component, AI will provide a clean, consistent snippet that perfectly matches your prompt. But here is the danger: LLMs are the ultimate yes-men. They will rarely push back and tell you that the feature you are building is fundamentally flawed, or that the architecture you are releasing is absolute rubbish.

This creates a dangerous divide. An already experienced developer can look at the generated code, question its validity, and make an informed judgement call on whether it is genuinely acceptable for the production environment. They have the hard-earned scars to know when a shortcut will result in technical debt.

Conversely, novice developers or non-technical managers wielding these tools can quickly fall victim to delusions of grandeur. Just because they can suddenly spin up a functioning web app or a complex API in an afternoon, they begin to believe they possess senior-level engineering prowess. However, software engineering is not just about stringing together functioning isolated components; it's about cohesive architecture, long-term maintainability, and understanding how a change in one area of an application effects the entire system.

When you blindly stitch together AI-generated code for months on end without that seasoned oversight, the results aren't going to be pretty. To quote a fellow developer I know, it becomes "AI slop". You eventually wake up to a codebase filled with inefficiencies, repetitive patterns, and short-sighted design choices. AI was consistent with the immediate prompt, but it failed entirely to maintain the long-term context of the project's evolution.

Conclusion

So, what is the solution? It certainly isn't a total retreat back to the analogue days of manual coding. The precedent Claude Code has set proves that AI is far too valuable a tool to discard. The answer lies in finding a fine balance.

We must stop treating AI as an outsourced developer that writes our code from start to finish, and start treating it as a brilliant, if occasionally short-sighted, peer reviewer. When AI offers a solution, we shouldn't just hit "accept" and move on. We need to dissect it. We must judge if its suggestion truly fits the broader architecture, learn from the new techniques it introduces, and actively verify its logic.

This approach keeps you engaged in the "why" and "how" of the code rather than just the "what". AI cannot be allowed to act as a substitute for human reasoning. It is there to assist, not to take the steering wheel completely.

We are the engineers; AI is the co-pilot. By finding this balance, we maintain was is required to actually learn and grow, ensuring we build software that stands the test of time without losing the joy of the craft itself.

In my previous post discussing my foray into the world of AI, I mentioned working on a personal project called "Stockmantics". But what exactly is Stockmantics, and why did I decide to build it?

Stockmantics started because I needed a project where I could apply my AI knowledge to a real-world problem. In the end, I didn't have to look further than my own hobbies.

Aside from coding, I’ve become heavily invested (pun intended) in the stock market. It all started shortly after COVID, when there was so much buzz online about people putting money into companies and index funds. Seeing the returns made by those who invested at the right time (during the lockdown of March 2020) opened my eyes to a potential new income stream. I didn't want to miss out on the fun, so I decided to learn the ropes of an area I knew nothing about. I just didn't expect it to turn into a full-time hobby.

However, unlike most hobbies, I considered this one fraught with danger; one must err on the side of caution. After all, real money is at stake, and acting foolhardy or investing incorrectly can lead to significant losses.

The Requirement

When I became more confident in my investment strategy and the type of trader I wanted to be, I found one aspect consistently time-consuming: finding an easy-to-read daily digest in one place. I was tired of hopping from website to website or subscribing to endless newsletters just to get a clear picture.

So, with the help of AI, I decided to build a tool that would do this for me, and Stockmantics was born. My requirements were as follows:

• Market Snapshot: A quick look at key indices (S&P 500, FTSE 100, NASDAQ, Commodities, etc.).
• Daily Summary: A single, concise sentence summarising what happened that day.
• Global News: Key events from the USA, Europe, and Asia.
• Crypto Updates: High-level developments in cryptocurrency, focusing on the majors.
• Investor Action: A conclusion based on the day's news, suggesting what an investor should look out for.
• Smart Glossary: Tooltipped definitions for stock market, investment, and economic terms to assist novice investors (and provide a constant refresher for myself).
• Social-Media Integration: Automatic posting to X, highlighting key stories from the day's article.

My philosophy for this personal project is simple: if it assists my own needs, that is a big win in itself. If someone else finds my method of digesting the day's financial news useful, that will be the icing on the cake. I decided early on that the success of Stockmantics would not be measured by visitor numbers or X followers, but by what I learnt during the development process and whether it truly works for me.

Application Architecture

The application architecture is based on the following Microsoft technologies:

ASP.NET Core Razor Pages

The website is a relatively small and simple application that consisted of the following pages:

1. Homepage
2. Article Listing
3. Article
4. Generic Content (for About/Terms/Disclaimer pages)

A CMS wasn't needed as all content and data would be served from Azure Storage Tables. All there is from a content-management perspective is an authenticated "Article Management" area, where content generated by Gemini could be overridden when required.

Azure Storage Tables

I actively decided to use Azure Storage Tables over an SQL database to store all of the Stockmantics data as there was no relational element between each table. It also provided a lower cost alternative and quicker route to development.

List of tables:

• Article
• MarketSnapshot
• SocialShare
• StockmarketGlossary
• AppSetting

Azure Blob

For images that may be used in article content.

Azure Functions

All the grunt work getting the data is done by Timer Triggered Azure Functions that would fire shortly after the US Markets open (around midday GMT) in order to get the most up-to-date goings on in the market.

A breakdown of the Azure Functions are as follows:

• Generate News Article - queries stock market API's and news feeds to send to the Gemini API to construct an article tailored to my requirements. It is then stored in the Article table with related attributes and additional meta data suited to be served in a webpage.
• Generate Social Posts - extracts 10 key facts from the generated news article to be transformed into tweets. The days generated tweets are stored until pushed to social media platforms.
• Market Snapshot - uses the Yahoo Finance API to return the market price and percentage change for the core market indices. These values are then passed to the Gemini APIs "Grounding with Google Search" to provide sentiment and the reasons behind the change in price.
• Post To X - publishes a tweet every 15 minutes.
• Post To Bluesky - publishes a post every 15 minutes.

The Chosen AI Engine

It was always going to be a choice between Google Gemini and OpenAI. I was already familiar with both LLMs (Large Language Models), having casually thrown stock market queries at them—among other things—long before this project was even a glint in my eye. Ultimately, my decision hinged on two key factors:

1. API: The ease of use and the reliability of the endpoints in returning structured data.
2. Cost Factor: Being unfamiliar with the specific pricing structures of LLMs, I needed to estimate the cost per API call and project my monthly expenditure based on token usage. The OpenAI GPT API Pricing Calculator provided an excellent breakdown of costs across all major AI providers.

I concluded that Google Gemini was the best fit for Stockmantics, primarily because the model I intended to use (gemini-2.5-flash) offered the most competitive pricing. The cost for one million input and output tokens works out to approximately $0.37, compared to OpenAI's $2.00.

Furthermore, I felt that Gemini held a slight edge over OpenAI. They might have been late to the AI party, but they have certainly made up for lost time with impressive speed. It also had a card up its sleeve that I only discovered during development: Grounding with Google Search. This feature allows the model to access real-time information from the web, ensuring that the data returned is current rather than limited to a training cut-off date.

Misjudging the Machine: Data is King!

I initially was under the impression that I could simply ask the likes of OpenAI or Gemini to collate the day's stock market news, which I could then format to my liking. However, this proved to be a mistake. When dealing with fast-moving financial news, I found the results hit-and-miss. The models would frequently return information that was out of date or cite entirely incorrect market prices (even when using Grounding with Google Search).

At this point, I realised I needed to take a step back and reassess my approach. It became clear that without a reliable, accurate data feed, this application would be of no use to man nor beast.

The solution had to start with raw data, which the LLM could then use as its base to expand upon. For this, I found pulling financial data available through the likes of Yahoo Finance feeds to be invaluable, amongst other finance-related news feeds.

Lengthy Vetting Period

The transition from a proof-of-concept to the final version of Stockmantics required a lengthy vetting period, which continued weeks after releasing to live. The raw output from the LLM was rarely perfect on the first try, leading to a many iteratation of refinement. My focus was on four key areas:

• Structure & Flow: Tweaking the system instructions to ensure the output was digestible, preventing the model from generating dense, unreadable paragraphs.
• Sector Balance: Ensuring the article provided a holistic view of the market, rather than fixating solely on volatile tech stocks or the "Magnificent Seven".
• Glossary Precision: Fine-tuning the tooltips to provide definitions that were accessible to novices without losing technical accuracy.
• Geopolitical Neutrality: Ensuring that reports on world affairs, which often drive market sentiment were delivered with an objective and balanced tone.

What I learnt from this process is that while anyone can write a basic AI prompt, getting the granular nuances right takes a significant amount of time. It is less about coding and more about the art of communication; you have to learn how to speak the model's language to get the consistent, high-quality output you need. Even now, I find myself still making ongoing tweaks for further improvement.

If you compare the very first article published against one the more recent, I am hoping a vast difference will be noticed.

Breakdown of Costs

One of my main priorities was to keep the running costs on this project tight and I think things ended up being quite good on value. Here is a monthly breakdown:

1. Website and Domain: £6.25
2. Azure Services (Functions/Blob Storage/Tables): £1.10
3. Google Gemini API: £4.00

So we're looking at around £11.35 in total monthly costs. Not bad. Google Gemini costs will be the only item that I expect to fluctute based on the varied number of tokens utilised for each daily article.

NOTE: Google Gemini and Azure services are only used weekdays for when the stock markets are open. So the costs are based on a 5 day week.

Conclusion

I am unsure what the long-term future holds for Stockmantics. Its lifespan ultimately depends on ongoing costs, maintenance effort, and whether I continue to find it useful for my own needs. However, for now, it serves a valuable purpose beyond just financial news: I have a robust, live application that acts as the perfect test bed for experimenting with new AI features and expanding my technical skillset.

Fortunately, thanks to various architectural decisions and efficiency improvements, the running costs are currently sustainable, and the site itself is very low maintenance—touch wood! I foresee that further development will only be required if the external APIs change. I have already paid for a years worth of web hosting until October 2026 and will reassess things closer to that date.

If you got this far, thank you for taking the time to read through the development process. If you are interested in seeing the final result, you can find all the links to Stockmantics below:

• Website - www.stockmantics.com
• X - Stockmantics
• Bluesky - stockmantics.com

What happens when we remove the human from a fact or a piece of information? Does it change how we perceive it? This thought came to mind when I was questioning if community-based sites, such as Stackoverflow are still relevant and made an open-ended remark that Generative AI has now starved us of knowing the person behind the knowledge.

Historically, we accept knowledge through some form of testimony. We will only believe in something based on what a person has told us. We evaluate their character, their knowledge and most importantly, their honesty. With AI, there is no "person" to trust. You cannot evaluate the AI's moral character or life experience because it has none.

To demonstrate this point, let's take the following statement about the US economy:

The stock market is the highest it's ever been. We have the greatest economy in the history of our country.

If you heard this from Donald Trump (the above statement has been said multiple times by him), you would likely question it immediately. We are familiar with his rhetorical style in how he often bends the truth or prioritises hyperbole over precision. Our scepticism is triggered by the source.

However, if you asked a financial analyst, you would get a more nuanced response:

While the market did hit record numbers (which happens naturally due to inflation), the rate of growth was not actually the 'greatest in history'. At the three-year mark, the market was up roughly 45% under Trump, compared to 53% under Obama and 57% under Clinton.

When we remove the human source, we lose this critical context. By stripping away the "who", we put the accuracy of the "what" in jeopardy. AI operates by taking the insights that required years of research and lived experience, strips them of their author, and repackages them only to regurgitate them with its own bias for our instant consumption. I rarely see the likes of ChatGPT or Gemini offer true attribution to the human behind the data for our own vetting.

I am far too aware of this from my own experience in building one of my own projects with AI focusing on the stock market and economy, where the data can be subjective and context-dependent. An example of this is when trying to provide the reasoning behind changes in key indices and commodities. The reasoning behind a change in value often hides a dozen competing narratives. When I built my application, I realised that if the AI chooses one narrative over another without telling me why or who championed it, it isn't just summarising the truth; it is effectively editing it.

Now, I don't want this post to come across negative towards AI, as it would pretty hypocritical after my glowing take on how I use the technology detailed in my previous post, it has just made me more conscious that even though knowledge it presents doesn't necessarily lack meaning, but it might lack soul. We get the answer, but we miss the human condition that made the answer necessary in the first place.

We have to acknowledge that AI is an incredible tool for gathering information, but it should be the starting point, not the finish. Use it to broaden your search, but go to people to deepen your understanding.

The title of this post isn't just a great line from Inception; it's a directive. Eames telling Arthur to expand their constructed reality beyond mere imitation and take bigger risks has been replaying in the back of my mind lately. It felt like the only appropriate way to break the radio silence after such a long hiatus and offer a glimpse into my current mindset. While I haven't been navigating multiple levels of a subconscious dream state, this past year has been about breaking free from self-imposed limitations. I've been pushing beyond my day-to-day coding endeavors to invest time into the very thing dominating our headlines: Artificial Intelligence!

It is a technology moving at such breakneck speed that you can't just dip a toe in; you have to dive in headfirst and swim, trusting that you'll emerge on the other side a wiser man. Failing to observe the shift in an industry like mine, in my view, is career suicide. With platforms and services releasing their own form of AI tools—some I deem more successful than others—I needed to find my own way in. As programmers, we can no longer afford the luxury of being so tunnel-visioned, clinging rigidly to our area of expertise while the landscape changes around us.

The thought of getting any footing into the world of AI filled me with dread. This could be down to setting the bar of expectation too high. I knew I was never going to be the type of person to build some deep learning AI engine from scratch, as you really need the "street smarts" of an AI Engineer to do that. Instead, learning to use AI tools and frameworks already readily available would give me the step up I needed, such as Machine Learning and APIs provided by ChatGPT and Gemini.

The Journey To Discovery

My journey began not with complex neural networks, but with the fundamentals of machine learning (via ML.NET). It was a learning curve, requiring me to rethink how I approached problem-solving. But as the concepts started to click, the potential for a specific use case suddenly became undeniable. I started small, experimenting with a simple concept that could be of tangible value, where I could predict future pricing of used cars based on historical data and their individual attributes.

Not too far along from this, I started working on my very own side-project in another area I am very passionate about: stocks and trading. I developed a website called Stockmantics that would take in the day's stock and trading news to produce daily digest in a format that was beneficial to me. My own one-stop shop for the day's trading news, without having to read many different newsletters as I had done previously. I used AI as a way to assist in my own needs that could also help others. It's a beast of a project that I am incredibly proud of, and I plan to do a write-up on it next year. But for now, suffice it to say that it taught me more about the practical pipelines of AI than any tutorial ever could.

One of the final AI projects I worked on at the tail end of the year was a proof-of-concept that revolved around vision search. I wanted to see if I could build a system capable of scanning a client's database to find visually similar items based on nothing but an uploaded image, with the ability to detect what the image consisted of. The addition of metadata attribution working alongside the image search resulted in accurate results that surpassed my own expectations.

If Asimov had his Three Laws to govern the behaviour of robots, I had my three specific applications, each being a critical stepping stone that would shape my understanding as to where I could integrate AI and the future possibilities—endless? Rather than just being the end user, I was building something of my own creation. I was able to see AI through a different perspective, which resulted in a newfound appreciation. It ended up being a really rewarding experience that has been far from what I am normally used to developing, and this is just the start.

Final Thoughts

I've come to view AI not as a competitor, or a full human replacement, but as a tireless, low-cost assistant ready to help take the smallest seed of an idea and grow it into a tangible reality, at a speed I never thought possible. It bridges the gap between theory and fruition, allowing me to truly dream a little bigger.