I built a crypto trading bot, pointed it at a real money account, and lost. I didn’t tell many people. I just quietly turned it off and moved on. That was the sensible thing to do. But something dropped last week that made me want to try again, and this time I’m doing it in public.

Let me back up.

The Original Plan Was Not Complicated

I am an IT professional. I spend my days around systems, automation, and infrastructure. When AI coding tools started getting genuinely good, the idea of building a trading bot felt less like a fantasy and more like a weekend project. So I built one.

The logic was straightforward. Run scheduled scans across a handful of crypto assets. Pull in historical trade data and combine it with market sentiment from a few free APIs. Generate an evaluation score for each asset. If the score crossed a threshold, the bot would buy or sell automatically. No emotion. No staying up watching charts at 2am. Just clean, systematic execution.

I called it tbot. I even gave it a home at tbot.augustwheel.com.

And honestly? The idea itself was not bad. It was the execution, the data, and the confidence that caused the problems.

Also Read:

• 12 AI Projects You Can Actually Build Without a Technical Background
• Project 1: Personal News & Keyword Alert System
• Everyone’s Talking About Clawdbot. So I Installed It. (And I’m Impressed)

What Actually Went Wrong

The first issue was the data. I was pulling sentiment signals from free API sources, which sounds fine until you realise what free actually means at scale. Limited calls. Delayed data. Gaps in coverage. When your bot is supposed to react to market conditions in near real time, stale sentiment data is not a minor inconvenience. It is the whole problem.

The second issue was that I had not properly backtested the strategy before connecting real money. I had done enough testing to feel confident, which is a very different thing from doing enough testing to be confident. The bot looked reasonable in my informal checks. Live markets are a different environment entirely.

The third issue, and I will be honest about this, is that I went in hoping to make money. Not as a long term experiment. Not purely for the learning. I genuinely thought this could generate some passive income. That mindset made me move faster than I should have, skip steps I should have taken seriously, and use real funds before the system had earned that trust.

A few hundred dollars later, I turned it off.

I Was Not Alone in This

Here is the thing I have learned since then. This experience is incredibly common, even among people who know what they are doing.

One developer documented building almost the same system I built, using moving average signals combined with GPT-powered sentiment analysis from social media. Early results were promising. Then false positives from sarcasm on Twitter, latency issues, and a model that had been overfit to historical data combined to produce a losing week that erased the early gains. The lesson they drew was that AI is powerful but markets are humbling.

Another developer spent four years running failed experiments before finally achieving consistent results, and the breakthrough only came after months of rigorous backtesting before a single dollar of real money was deployed.

The pattern across almost every honest account of building AI trading systems is the same. The demo works. The backtest looks good. Live markets expose every assumption you made.

One analysis of retail trading bots put it plainly: most barely break even, fees eat into gains, market noise throws off signals, and professional algorithms leave smaller players behind. Anyone promising otherwise is selling something.

I was selling myself the dream. The market charged me tuition.

So Why Try Again?

Last week Anthropic, the company behind Claude, released a document called The Complete Guide to Building Skills for Claude. Twitter picked it up and immediately misrepresented it as a ready-made trading bot that earns $300 to $1,500 a day. That claim is not accurate. I read the actual document.

What it actually describes is a framework for building Claude Skills, which are structured instruction sets that teach Claude how to handle specific, repeatable workflows consistently. I went back and read every page. There is no prediction market trading bot anywhere in the document. The screenshots circulating on Twitter appear to have been taken from a completely separate source and attached to the Anthropic document to make the post go viral. The real document contains examples like sprint planning tools, design-to-code handoffs, and payment compliance workflows. Useful and genuinely interesting, but not a trading bot.

Here is why that matters for this project. The Claude Skills framework is still exactly what I want to use for the tbot rebuild, but for the right reasons rather than a fabricated one. The framework allows you to encode a decision-making methodology directly into Claude, one that can apply contextual reasoning consistently across every scan cycle without needing to be re-instructed each time.

That is genuinely different from what I was doing before.

My original bot was a script making decisions based on rules I had written. If the score is above X, buy. If it falls below Y, sell. The logic was mine, limited by what I could think to code. What the Claude Skills framework enables is something closer to encoding a decision-making methodology, one that can incorporate multiple data sources, apply contextual reasoning, and flag when conditions fall outside expected parameters.

It does not make the market more predictable. No AI does that. But it does make the system more intelligent about when to act and when to stay still, which is exactly where my original bot failed.

What Happens Next

I am rebuilding tbot. You can follow the whole thing at tbot.augustwheel.com.

This time the rules are different. No real money until the system has proven itself over an extended paper trading period. Paper trading means real market data, real signals, real execution logic, but no actual funds at risk. If I cannot make it work in simulation, I have no business running it live.

I will be using Anthropic’s Claude Skills framework as the foundation for the decision layer. I will document every build decision, every failure, every adjustment. The losses from the first attempt will stay in the story because they are part of it.

I am not promising this will be profitable. I genuinely do not know if it will be. What I do know is that the tools available now are meaningfully better than what I had before, and the only way to find out what they can actually do is to build something real with them and show you the results.

That is the experiment. Come watch.

---

Follow the rebuild at tbot.augustwheel.com. If you want to stay updated as each build post drops, the August Wheel newsletter is the best place to catch them.