The Spell Book

Dacia Ventures
50 spells from 49 books. One-word triggers for complete thinking frameworks.
🔍
50
Spells
5
S-Tier
18
A-Tier
16
B-Tier
11
C-Tier
S
S-Tier - Use Constantly
A
A-Tier - High Frequency
B
B-Tier - Situational
C
C-Tier - Utility

The Bestiary

Describe your problem. The Bestiary will diagnose the monster and prescribe the spells.
We ship features but nothing moves the needle System works in testing but breaks in production Team is stuck and can't make progress We keep fixing bugs but more appear AI outputs are confident but wrong
Describe what's broken, or click an example above.
The Bestiary will find the monsters lurking in your problem.

Combat Chains

Pre-built spell sequences. Click any spell in the chain to jump to its definition.

Install Into Any System

Copy these into any LLM's system prompt, project instructions, or CLAUDE.md. Instant spell casting.

Full Install

All 50 spells with casting protocols, chains, and bestiary lookups.

100% effectiveness

Core Spells Only

S-Tier + A-Tier spells. The 22 you'll use 80% of the time.

80% effectiveness - fits smaller context windows

Trigger List

Just spell names and one-line descriptions. Minimal footprint.

70% effectiveness - LLM fills in the gaps

The Bad News

Why this spell book exists. Why you need to move NOW.

Claude Turned $1,000 Into $14,000 on Polymarket in 48 Hours

March 2026. A solo developer deployed a 274-agent AI trading terminal on Polymarket (a prediction market). The system used an ensemble of three frontier LLMs — GPT-4o (40% weight), Claude 3.5 Sonnet (35%), and Gemini 1.5 Pro (25%) — making probabilistic bets on real-world events.

The result: $1,000 → $14,000 in 48 hours. A 1,300% return. No human intervention after deployment.

How It Works — The Architecture

Ensemble Forecasting: Each LLM independently generates a probability estimate for market outcomes. These are combined using weighted averaging (GPT-4o 40%, Claude 35%, Gemini 25%) to produce a final probability. This is exactly how ensemble methods work in ML — multiple weak learners combining into a strong one. See: ENSEMBLE spell.

Fractional Kelly Criterion: The system uses Kelly sizing with 7 multipliers to determine bet size. Kelly criterion maximizes long-run growth rate by betting proportional to edge. The fractional variant (typically 0.25x–0.5x Kelly) reduces variance at the cost of some expected return. The 7 multipliers include: edge size, confidence level, market liquidity, time to resolution, correlation with existing positions, drawdown limits, and portfolio concentration.

Bayesian Probability Updates: As new information arrives, the system updates its probability estimates using Bayes' theorem — prior belief × new evidence = updated belief. Each LLM re-evaluates, and the ensemble re-weights. This is continuous learning in production. See: DRIFT spell.

Risk Management: 15+ automated risk checks before any trade executes. Includes position sizing limits, correlation checks, drawdown circuit breakers, and liquidity validation. The system has 4 execution strategies it selects from based on market conditions.

The Spells It's Casting

This trading bot is literally casting spells from this book:

ENSEMBLE — 3 LLMs weighted-averaging predictions.
DRIFT — Bayesian updates when reality shifts from priors.
RED TEAM — 15 risk checks adversarially testing every trade.
LOSS FUNCTION — Kelly criterion defining what "winning" means mathematically.
TEMPERATURE — Balancing exploration (new markets) vs exploitation (known edges).
REGULARIZE — Position limits preventing overfit to any single bet.

The future isn't humans using AI tools. It's AI systems using AI reasoning patterns autonomously. The Spell Book is your operating system for that world.

Citations & Sources

Primary source: Polymarket AI Trading Bot — open-source repository (GitHub, March 2026). 274-agent terminal architecture with ensemble LLM forecasting.
Kelly Criterion: Kelly, J.L. (1956). "A New Interpretation of Information Rate." Bell System Technical Journal. Extended by Thorp (2006) for fractional Kelly with multiple simultaneous positions.
Ensemble Methods: Dietterich, T.G. (2000). "Ensemble Methods in Machine Learning." Multiple Classifier Systems. Weighted averaging as implemented: GPT-4o 40%, Claude 3.5 Sonnet 35%, Gemini 1.5 Pro 25%.
Bayesian Updating: Jaynes, E.T. (2003). "Probability Theory: The Logic of Science." Prior × Likelihood = Posterior, applied continuously to market probability estimates.
Prediction Markets: Arrow, K.J. et al. (2008). "The Promise of Prediction Markets." Science. Polymarket as the execution venue for AI-generated forecasts.

How This Works

Spells are single-word triggers that activate complete reasoning frameworks. Say "SWOT this" and your agent runs a three-pass strategic analysis. One word, full framework.

Monsters are problems you'll hit. Each one tells you what's wrong and which spell fixes it. Counter-spell tags are clickable - they jump to the spell definition.

Chains are pre-built spell sequences for common situations. Every spell in the chain is clickable.

Install lets you copy spell definitions into any LLM system. Full install = 100% casting accuracy. Trigger list = 70% (the LLM already knows SWOT, but not YOUR 3-pass protocol).

The key: These aren't technical jargon. "Loss function" = define what you're optimizing. "Backprop" = trace failure to root cause. "Temperature" = how random should your thinking be? You already use these patterns. Now they have names.

50 Spells

From 49 AI/ML books. Each one a complete thinking pattern.

💀

20 Monsters

Problems you'll face. Click counter-spells to learn them.

10 Chains

Spell combos. Click any spell to jump to it.

📦

Install

Copy into any LLM. Instant spell casting anywhere.