staskh/trading_skills

A Claude advisor for options traders: 25 skills in SKILL.md format under .claude/skills/, each teaching the agent to run Python scripts against market data. MIT, 324★, 73 forks, 133 commits (2026-08-06). Coverage runs quotes, option chains, price history, fundamentals, news and sentiment, earnings calendars and insider filings; then indicators (RSI, MACD, Bollinger, SMA/EMA, ATR, ADX), Greeks, spread analysis and risk metrics (Sharpe, VaR, Piotroski F-score); then scanners for bullish trend, PMCC candidates and unusual options activity; then Interactive Brokers portfolio reads and PDF report generation.

Data comes from Yahoo Finance (free, roughly 15 minutes delayed), optionally the Massive API (ex-Polygon) for institutional options flow, and IB's TWS/Gateway for live portfolio state. Python 3.12 and uv. There is also an MCP server variant exposing 32 tools, which the README says runs on the free Claude tier — the skills version wants Claude Code or Cursor.

The second skillpack, and what it does to the first reading

superior-skills gave this spoke agent-deployed-trading, and that page drew a commercial conclusion from a single case: publish an open skills pack, keep the backtester, the data and the execution behind an API key, and the pack becomes a customer-acquisition channel. This repo is the control case that claim needed. It is also MIT, also an agent-skills pack for trading, and there is no platform behind it — Yahoo is free, IB is the user's own brokerage account, and the one paid dependency (Massive, for whale-hunting) is optional and is not the author's product. Nobody earns when you deploy.

So the customer-acquisition shape is one arrangement rather than the pattern, and the honest revision is narrower: an open trading skillpack tells you nothing about incentives until you ask what it needs a key for. Ask that of these two and you get opposite answers.

No performance claims, because there is nothing to claim them with

The sharper contrast is about evidence. superior-skills labelled a strategy "validated" on six trades, which is the specimen backtest-overfitting and chan-algorithmic-trading had just prepared the spoke to recognise. This repo makes no performance claim at all — and the reason is structural, not modest. It has no backtester and no strategies. It computes indicators, screens for setups, and prices Greeks; whether a PMCC screen has ever made money is not a question the repository is arranged to answer.

That removes the failure mode superior-skills demonstrated and installs a different one. The draft → backtest → review → deploy pipeline that Superior at least advertised is absent here, so the chain runs data → indicator → the trader's judgement → optionally an order. Nothing in the repository sits between a scanner hit and a position. A screen that surfaces candidates without a validation layer is the backtesting problem relocated rather than solved: instead of a backtest nobody can audit, there is no backtest, and the agent's fluency about RSI and Greeks is doing the persuading.

Where the human check actually sits

agent-deployed-trading named the worry that the last human control is an approval prompt. This repo answers it differently and the difference is worth recording precisely: IB access is read-only by default, and order placement requires switching the integration to read-write.

That is a real boundary, and it is not a per-trade one. It converts the human decision into a single configuration change made once, in advance, covering every order that follows. Better than Superior's custodial "no-key trading" in that the default is safe and the account stays the user's own; not a substitute for review, because after the switch the approval-prompt question returns unchanged. The spoke's two observations still apply — a trade has no diff, and a kill switch runs downstream of the mistake.

The delayed-data problem nobody flags

The README's own caveat, that Yahoo quotes lag ~15 minutes, is stated as a data-freshness footnote and is more than that for this specific application. options-greeks, spread analysis and unusual-options-activity scanning are the parts of options trading most sensitive to current price; delta and gamma computed off a quarter-hour-old underlying describe a position that has moved. The delay is the smaller of its two data problems: information-in-option-volume shows that the option-volume signal a public feed can reconstruct at all has no predictive power once the private open/close flag is in the same regression. For screening and study this is tolerable. The repository does not distinguish the uses, and its most prominent feature list mixes them.

Tier. T3 at ingest, superseded 2026-08-20 → T1 — a third-party repository's own README, self-described, with no external review and no measured claims. No explicit risk disclaimer, though it names its audience as retail self-directed traders.

Related

agent-deployed-trading · superior-skills · backtesting · backtest-overfitting · erik-smolinski (the spoke's other options practitioner, working the same free-data stack by hand) · synthesis · cross-spoke: ../agentic-tooling-wiki owns the skills standard and the packaging

Re-tiered 2026-08-20 → T1. Under the corpus rule adopted this day (../QUALITY.md → tier by the claim, not the URL), this page rests on the pack's inventory — twenty-five SKILL.md skills and the data, indicator and Greeks coverage each one wraps. A project is the primary source for its own design, and the earlier grade tiered the URL rather than the claim.