How PassiveBot AI actually works

Most crypto trading bots are black boxes. You deposit money, hope for the best, and get a vague dashboard with green numbers until the day it blows up. PassiveBot is different. We're building it in public, and this page explains the exact architecture — every system, every check, every safeguard.

The core problem with most crypto bots

Automated crypto trading bots fail for two main reasons: they trade in conditions they shouldn't, and they size positions incorrectly. A bot with a 60% win rate can still lose money if the average loss is larger than the average win. PassiveBot was designed from the ground up to address both of these problems.

Before any trade fires, PassiveBot runs 15 independent checks. If any single check fails, the trade is skipped. This sounds conservative — and it is. The bot will often sit on its hands for days. That is the point. In a market like April 2026, with Fear and Greed at 21, the correct action is usually to do nothing.

The triple AI consensus system

PassiveBot uses three separate AI models from three different AI labs: Claude (Anthropic), DeepSeek, and Gemini Flash (Google). Each model is given a different persona and asked independently whether the trade setup is worth taking. All three must agree before a trade fires.

This is not a gimmick. Running three models from different labs means the consensus is genuinely independent — there is no shared training data or shared bias. If one model has a blind spot, the other two will catch it. The AI acts as a veto layer, not a trigger. The technical setup must already pass 15 checks before the AI is even consulted — meaning the AI only gets asked about genuinely promising setups, keeping LLM costs to $1–3 per month.

The 15-check filter

Every potential trade must pass all 15 checks before it is considered. These checks cover trend alignment across multiple timeframes, momentum, volatility, market regime, Fear and Greed index gating, funding rate environment, liquidity conditions, macro calendar events, and more. No trade fires on a hunch or a single indicator. Every check must pass.

The self-improving ML loop

PassiveBot uses an XGBoost meta-learner that observes every trade and every skipped signal. After 30 trades of data, the model begins proposing parameter adjustments — slightly wider stops, different entry thresholds, adjusted position sizing. Every proposal is reviewed by the human operator on the Sunday weekly review and approved or rejected. The bot gets smarter every week, but the human stays in control of every change.

Safety systems

The weekly review

Every Sunday, paste the weekly review output into any AI assistant for a 30-minute session — review trades, backtest improvements, approve any changes. It summarises the week's scans, trades taken, signals skipped, heartbeat status, kill switch events, and any ML proposals from the XGBoost loop. The entire review takes about 30 minutes. You read the report, approve or reject any ML proposals, and go back to your weekend.

Your daily routine

Once set up, PassiveBot runs automatically. Your involvement is about 5 minutes each morning and 30 minutes on Sunday.

Every morning — 5 minutes

Open any AI assistant (ChatGPT, Claude, Gemini — whatever you use). Paste the included prompts/2-DAILY-AUDIT.md file along with your overnight terminal output. The AI reviews what traded, checks the system is healthy, flags anything unusual, and suggests one improvement if there is anything worth changing. Five minutes. Done.

Every Sunday — 30 minutes

Review the week's trades. Paste prompts/3-IMPROVE.md into any AI assistant with any idea you have — a filter to test, a coin to remove, a parameter to change. The AI backtests it against 3 years of data and gives you a clear yes or no with the numbers. You approve or reject. Nothing changes without your say-so.

Setup — 20 minutes, no coding

Paste the included prompts/1-SETUP.md file into any AI assistant. Answer a few questions about your machine. The AI configures your paths, installs dependencies, sets up your scheduler, runs the first audit, and sends a test Telegram message. Works on Mac, Linux VPS, and Windows.

If anything ever breaks, paste the error into any AI assistant. This system was built using AI. Every AI assistant understands the architecture and can fix it.