Thread of Claude prompts for quant work

A viral social thread turned Claude into a mock quant desk, packaging 12 copy-and-paste prompts for forecasting, backtesting, portfolio construction and execution analysis. (threadreaderapp.com) The thread was posted by X user Nav Toor and archived by Thread Reader App as a 15-post thread dated February 16. Its opening claim said Claude could build machine-learning trading models “for free” and framed the prompts around roles at firms including Goldman Sachs, JPMorgan and Citadel. (threadreaderapp.com) The prompts themselves read like job briefs. One asks for a full time-series forecasting model with data cleaning, feature engineering, model comparison, backtesting, risk rules and Python-style implementation notes; another asks for a mean-reversion strategy with cointegration tests, Kelly sizing and drawdown limits. (threadreaderapp.com) Before the thread, the underlying workflow was already familiar on trading desks: quants turn market data into rules, test those rules on old data, then check whether the strategy survives realistic trading costs and risk limits. The United States Securities and Exchange Commission’s report on algorithmic trading describes institutional execution algorithms as tools that decide when, where and how to place orders. (sec.gov) That is where “market microstructure” comes in. The field studies how prices form inside the market’s plumbing — bid and ask quotes, order queues, trading venues, liquidity and the cost of getting filled — rather than just whether a stock went up or down. (coursera.org) The Jane Street-style angle in the thread points to speed and execution, not just prediction. Jane Street says its own systems are built for high throughput, low latency and strong reliability, and one of its engineering talks describes processing millions of messages per second for electronic trading. (janestreet.com, janestreet.com) Anthropic has spent the past two years pushing the same basic lesson from the model side: better prompts produce more accurate, more consistent and cheaper outputs. In a February 29, 2024 post, the company said users should give Claude clear instructions, specify output formats and ask it to work step by step. (anthropic.com) Anthropic also teaches prompt design as a structured skill, not a magic phrase. Its public tutorial breaks prompting into basics such as clear instructions, role assignment, separating data from instructions, step-by-step reasoning, examples and methods for reducing hallucinations. (github.com) The finance use case is no longer hypothetical on Anthropic’s own site. Claude’s current use-case library shows finance workflows such as updating a financial model after earnings, drafting a credit memo from spreads and statements, and validating reserve workbooks with external data connectors. (claude.com) What the viral thread adds is packaging. Instead of teaching prompt craft in the abstract, it wraps Claude in familiar Wall Street personas and asks for outputs that resemble research memos, trading playbooks and model specs. (threadreaderapp.com, github.com) The catch is that a polished prompt is not a live trading system. The thread asks Claude for “expected accuracy,” Sharpe ratios and win rates, but regulators and market-structure specialists treat execution costs, latency, liquidity and model risk as separate problems that still have to be measured against real data and real fills. (threadreaderapp.com, sec.gov, coursera.org) That leaves the thread as a snapshot of where generative artificial intelligence is landing in finance in 2026: less as an autonomous trader, and more as a fast draft machine for work that quants, engineers and risk teams still have to verify. (anthropic.com, claude.com, sec.gov)