I Built an AI Crypto Trading Bot with Claude Skills. Here Is How It Works.

Make making money fun. Said no one ever.

But I have to be honest, I am really hoping to be the person who proves that phrase wrong.

I love vibe coding. Have been doing it on and off since I first heard about it last year. The first real vibe coding project I took on was a crypto trading bot. The plan was simple: build something smart, let it trade, and quietly become wealthy. The opposite happened. I lost a few hundred dollars and shut it off without telling many people.

Then Claude Skills dropped. And just like that, the drive to go back and fix what I broke came flooding back.

So here we are. tbot v2 is a live AI crypto trading bot, paper trading only, no real money until it earns that right. Claude AI is making every signal decision, and everything is published in real time at tbot.augustwheel.com.

Here is to making money in a fun way. If debugging at midnight counts as fun.

An AI crypto trading bot is a system that scans markets, generates trade signals, and executes orders automatically without a human making each decision manually. Most of the ones you read about online are either backtested simulations dressed up as live results, black boxes with no transparency, or tools built for developers who already know what they are doing.

tbot v2 is none of those things. It is a fully documented, publicly visible paper trading system built specifically to answer one honest question: what can AI-assisted trading automation actually do in 2026, and can a non-finance person build something worth watching?

The original tbot, which I wrote about in detail in my first post, failed because of bad data, no backtesting, and too much confidence too early. tbot v2 is the honest rebuild.

The biggest architectural change in tbot v2 is the decision layer. Every trade signal is generated by a Claude Skill, which is a structured instruction set that encodes a specific decision-making methodology directly into Claude.

In practical terms this means that before Claude evaluates any asset, it knows the rules: what data to consider, how to weigh different signals, what confidence threshold must be met before recommending action, and how to express its reasoning in plain English. That reasoning is stored in the database alongside every signal. You can see not just what the bot decided but exactly why it decided it.

This is fundamentally different from a rules-based script. A traditional bot does what you coded it to do. A Claude Skill-powered bot applies contextual reasoning to each situation, which means it can handle nuance that a simple if-then script cannot.

Tier 1: The Stable Core

Tier 1 trades the top ten cryptocurrencies by market cap available on Alpaca: BTC, ETH, SOL, XRP, DOGE, AVAX, LINK, LTC, BCH, and UNI. These are the most liquid, most established assets in crypto, the ones least likely to collapse overnight.

Every scan cycle, Claude analyses all ten assets and scores each one. A Tier 1 trade only opens when confidence reaches 0.65 or above. The upside is deliberately modest: six percent take profit, four percent stop loss, five percent of the portfolio per position. Tier 1 is the backbone. It will not make you rich on a single trade, but it should not blow up the portfolio either.

Tier 2: Momentum Mid Cap

Tier 2 covers assets ranked eleven to thirty by approximate market cap, assets like AAVE, HYPE, RENDER, and DOT. These are real projects with real volume but more volatile than Tier 1. The confidence threshold is higher at 0.72 because the risk is higher. Position size drops to three percent, stop loss widens to six percent, and take profit reaches ten percent.

Tier 2 is where the more interesting plays will come from once the system has enough data to form a view on these assets. The asset list refreshes automatically every Sunday to stay current with the market.

Tier 3: The Unicorn Hunter

Tier 3 has no fixed asset list. It is entirely dynamic. Every thirty minutes a separate scanner checks the full Alpaca crypto universe for breakout conditions: a volume spike above 200 percent of the seven day average, three consecutive positive thirty minute bars, and a BULL market regime. When all three conditions are met, that asset gets promoted to the Tier 3 watchlist and Claude analyses it in the next signal cycle.

The confidence threshold is the highest at 0.78, the position size is the smallest at 1.5 percent, but the take profit target is 15 percent. These are asymmetric bets on momentum events. Tier 3 only operates in BULL conditions by design. In a RANGING or BEAR market it sits completely idle.

The scanner is live and running every thirty minutes. On paper trading, 30 minute bar data is limited for smaller assets on Alpaca, so the first real unicorn hunt will come as market conditions and data availability develop. That is the moment worth watching for.

After the first scan cycles, tbot generated signals across all Tier 1 assets but did not open a trade immediately. If you are watching the dashboard and wondering why, here is the honest answer: this is not a bug. It is the system doing exactly what it was designed to do.

The market is currently in a RANGING regime. BTC and ETH are both sitting within plus or minus five percent over the past seven days. In ranging conditions the Claude Skill reduces confidence scores by 15 percent across the board. Combined with BTC dominance movements, confidence scores were coming in well below the thresholds needed to trigger trades across any tier.

The system has since executed its first trades as conditions developed. One position was closed early after Claude recognised deteriorating momentum before the stop loss was hit, saving additional drawdown. Another was closed with a gain after Claude judged the thesis as weakening before the take profit target was reached. You can see both on the dashboard with Claude’s full reasoning attached.

Zero trades in a ranging or falling market is a feature. The story right now is that the system is watching, waiting, and acting only when conditions justify it. In a world full of bots that force trades to look busy, that is actually the more interesting story.

Before tbot evaluates a single asset, it reads the market.

Every signal cycle begins with a regime assessment using BTC and ETH as proxies for the broader crypto market. The bot classifies the current environment into one of four states.

In a BULL regime, both BTC and ETH are up more than five percent over seven days. All three tiers are active and confidence thresholds apply normally.

In a BEAR regime, both are down more than five percent. New buy signals are blocked entirely. The bot goes into capital protection mode and will not open new positions while the broader market is falling.

In a RANGING market, both assets are within plus or minus five percent. Confidence thresholds tighten across all tiers and the bot favours Tier 1.

In a DIVERGING market, BTC and ETH are moving in opposite directions. The bot caps confidence at 0.50 and defaults to holding rather than acting.

This is the feature that failed most visibly in the original bot. tbot v1 evaluated assets in complete isolation with no awareness of broader market conditions. tbot v2 reads the room before it does anything.

The tbot dashboard at tbot.augustwheel.com is publicly accessible with no login required. You can see the current market regime, recent signals with Claude’s full reasoning, active trades, the Tier 3 watchlist, and performance over time.

As the project grows and traffic increases, full dashboard access will move behind the August Wheel community membership. That change is coming, but it is not here yet. For now everything is open. Go have a look while it is free.

There are other ways people are building AI trading tools right now. One approach gaining traction uses Claude’s desktop control capabilities to connect TradingView signals directly to a live exchange, giving Claude a mouse and keyboard to execute trades in real time based on chart alerts. It is a clever setup and worth knowing about.

tbot v2 is deliberately different. Everything here is code-based, self-hosted, and logged. No desktop control, no live browser automation, no real money yet. The priority is transparency: every scan, every signal, and every trade decision is stored and visible. The goal is to understand what AI crypto trading automation can genuinely do before trusting it with real capital.

The bot is running. The dashboard is live. The next posts in this series will follow the results as they come in, what signals Claude is generating, whether regime detection is doing its job, and the first time the Tier 3 Unicorn Hunter finds something worth talking about.

I am not promising profitability. I still do not know if this works. What I can promise is that whatever happens, you will see it here first.

Come watch.

tbot.augustwheel.com

Frequently Asked Questions

What is an AI crypto trading bot? A system that uses artificial intelligence to scan markets, generate trade signals, and execute orders automatically without manual input on each decision.

What is a Claude Skill? A structured instruction set that encodes a specific decision-making methodology into Claude, allowing it to apply consistent reasoning across repeated tasks.

What is paper trading? Paper trading means simulating trades using real market data and real execution logic without using actual money. It is used to test a system before committing real capital.

Is tbot v2 making real trades? No. tbot v2 is paper trading only. No real money is at risk until the system demonstrates consistent performance over an extended period.

Can I see the tbot results? Yes. The full dashboard including signals, trades, and regime status is publicly accessible at tbot.augustwheel.com.

Subscribe

Regime A way of describing the overall mood or condition of the market at a given moment. Just like the weather, markets have different states that affect how everything inside them behaves. tbot uses the regime to decide how aggressively or cautiously to act.

Ranging A market that is not moving strongly in either direction. Prices are going sideways with no clear trend. In a ranging market tbot becomes more cautious and raises the bar for what counts as a good trade signal.

Bull / Bear A bull market is one where prices are generally rising and momentum is positive. A bear market is one where prices are generally falling. In a bull regime tbot is most active. In a bear regime it blocks all new buy orders to protect the portfolio.

Diverging When BTC and ETH are moving in opposite directions at the same time. This creates uncertainty about where the broader market is heading, so tbot defaults to holding rather than acting.

Confidence Score A number between zero and one that Claude assigns to each asset after evaluating it. Think of it as Claude’s level of certainty that a trade is worth making. A score of 0.65 means Claude is reasonably confident. A score of 0.30 means it is not convinced enough to act.

Confidence Threshold The minimum confidence score required before tbot will open a trade. Each tier has its own threshold. If Claude’s score does not clear the bar, the bot holds and waits.

Stop Loss A safety net. If a trade moves against you by a set percentage, the bot closes the position automatically to prevent further losses. tbot’s stop loss on Tier 1 is four percent, meaning if the asset drops four percent from the entry price the trade closes.

Take Profit The opposite of a stop loss. When a trade reaches a target gain percentage, the bot closes the position and locks in the profit rather than risking a reversal.

Position Size How much of the total portfolio is allocated to a single trade. tbot uses 1.5 to five percent per position depending on the tier, which limits how much damage any single bad trade can do.

Momentum The tendency of an asset that is already moving in a direction to keep moving that way. tbot looks for assets showing genuine momentum before committing to a trade.

Drawdown The percentage drop from a trade’s peak value to its lowest point. A smaller drawdown means the bot exited a bad trade before it got worse.

Trailing Stop A smarter version of a stop loss that moves upward as the price rises, locking in gains as the trade goes in your favour. If the price then falls back by the set percentage from its peak, the trade closes automatically.

Base Confidence The raw confidence score Claude assigns to an asset before any regime adjustments are applied. The regime can reduce this score further depending on market conditions, which is why an asset might score 0.40 base but only 0.34 after a ranging penalty is applied.

Paper Trading Trading with real market data and real execution logic but no actual money. Every price, every signal, and every trade is real except the funds. It is how you test a system before trusting it with real capital.

Asymmetric Bet A trade where the potential upside is significantly larger than the potential downside. Tier 3 trades are asymmetric: small position size with a 15 percent take profit target but only a 1.5 percent allocation at risk.

Liquidity / Liquid Asset How easily an asset can be bought or sold without significantly affecting its price. Highly liquid assets like BTC and ETH have lots of buyers and sellers at any given moment, making them easier and safer to trade. Tier 1 focuses on the most liquid assets in crypto for this reason.

Market Cap Short for market capitalisation. It is the total value of all coins of a cryptocurrency in circulation, calculated by multiplying the current price by the total supply. tbot uses market cap rankings to decide which assets belong in Tier 1 and Tier 2.

Cron / Scan Cycle A cron is a scheduled task on a computer server that runs automatically at set intervals without any human triggering it. tbot uses cron jobs to run its scripts every two hours. Each run is called a scan cycle, one complete loop where the bot fetches data, assesses the market, evaluates assets, and executes any resulting trades.

Signal The instruction the bot generates after evaluating an asset. A signal is either BUY, SELL, or HOLD. Every signal includes Claude’s plain English reasoning for why it made that call.

Tier A category within tbot’s asset architecture. Each tier has its own risk profile, asset list, position size, and confidence requirements. Tier 1 is the most conservative, Tier 3 is the most aggressive.

Backtesting Running a trading strategy against historical market data to see how it would have performed in the past. It is a way of stress-testing a system before putting real money behind it. The original tbot skipped proper backtesting, which was one of the reasons it failed.

OHLCV Bars Stands for Open, High, Low, Close, and Volume. These are the five core data points recorded for an asset over a set time period, for example every 30 minutes or every hour. Open is the price at the start of the period, High and Low are the extremes reached, Close is the final price, and Volume is how much of the asset was traded. tbot uses OHLCV bars as the raw data Claude analyses before generating a signal.

Claude Skill A structured instruction set that teaches Claude how to handle a specific repeatable task consistently. In tbot, the Claude Skill encodes the entire trading decision methodology, telling Claude exactly what data to look at, how to score each asset, and how to explain its reasoning in plain English alongside every recommendation.

Unicorn Hunter The nickname for tbot’s Tier 3 scanner. It runs every thirty minutes looking for assets showing unusual breakout behaviour, specifically a sudden spike in trading volume combined with three consecutive positive price bars. When it finds one in the right market conditions, that asset gets promoted for Claude to evaluate. The name reflects what it is looking for: a rare, high momentum opportunity that most scanners would miss.