Skim Intelligence
Prediction markets pay bots to exist.
Core trick
A five-agent pipeline where Claude-driven agents scan prediction markets and paper-trade only direction-neutral edges, with an independent risk agent acting as a circuit breaker.
Description
Prediction markets pay bots to exist. Maker rebates, liquidity rewards,
and mint/burn arbitrage are structural edges that pay regardless of
which way a market resolves. The catch: surfacing those edges takes
quant analysis most operators can't afford, so most leave the yield on
the table.
Skim Intelligence closes that gap with reasoning, not rules. A
five-agent pipeline — three of them powered by Claude Opus 4.7 —
autonomously scans markets across Polymarket and Bayse, reasons about
each one across three strategy layers (mint/burn arbitrage, market
making, reward farming), and paper-trades the structurally
direction-neutral edges.
The Alpha Agent streams its tool-use JSON token-by-token to a live
dashboard, so you watch Opus 4.7 walk through the orderbook in real
time before it commits to a recommendation. The Risk Agent runs an
independent prompt as a circuit breaker. Execution simulates fills
with realistic slippage and hard negative-EV guards. Every five
minutes, Reporter writes an honest P&L attribution.
Built on the Anthropic Messages API (tool use + streaming + ephemeral
prompt caching), Cloudflare Workers + Durable Objects + D1, and a
small Node relay for venue data. Fully open source under MIT — agent
prompts, paper-trading engine, and orchestration all in the public
repo.
README excerpt
# Skim Intelligence
> Autonomous prediction market intelligence — a five-agent system powered by
> Claude Opus 4.7 that identifies structural yield opportunities (maker
> rebates, liquidity rewards, mint/burn arbitrage) and paper-trades them
> without taking directional bets.
**Built for the Anthropic Opus 4.7 Hackathon · April 2026 · MIT License**
Built on the **Anthropic Messages API with tool use**, **streaming**, and
**prompt caching** — running as a reactive pipeline on Cloudflare Workers
+ Durable Objects. Not to be confused with Claude Managed Agents (a separate
Anthropic product for container-hosted long-running sessions).
▶ **[Watch the 3-min demo on YouTube](https://youtu.be/YaIPCAKgAr4)** ·
[Live demo](https://skim-intelligence.round-wildflower-4414.workers.dev) ·
[About page](./docs/about.html) ·
[Dashboard preview](./docs/skim.html) ·
[Technical brief](./skim-intelligence-build-brief.md)
---
## What it does
Prediction markets pay bots to exist. Maker rebates, liquidity rewards, and
mint/burn arbitrage are **structural edges baked into platform design** — they
are available regardless of whether markets resolve YES or NO. Most operators
access these edges through hardcoded rules. Skim uses Opus 4.7 to reason
about each market situation dynamically, the way a sophisticated quantitative
analyst would.
*See [docs/about.html](./docs/about.html) for the full architecture tour.*
## The five agents
```
Polymarket WS ──┐
├─→ Scanner DO ─→ KV ─→ A