Claude Day Trader Skill: What the 11-Command Pack Actually Does

On 20 October 2026 a repository called day-trader-skill went up on GitHub under the account Jakeschincariol, and by 23 October it had 637 stars on GitHub, read 23 October 2026. In the same window, Google Trends recorded "claude trader skill" and "github day trader skill" as breakout searches worldwide over the past seven days, read on 23 October 2026. The thing people are starring is not a trading bot. It is a set of eleven instruction files, eleven small Python tools, and a folder where each step leaves a note for the next one.

That distinction matters more than it sounds. A bot takes an action; this pack writes files into a trader/ directory and waits. The repository's README states the rule that governs everything else: "No trade goes through until you say yes. Nothing in this pack can place an order on its own. Paper trading is the default." Nothing below changes that, and nothing below is investment advice. This is MIT-licensed software that reads public data and does arithmetic, and the person who wrote it says as much in the README's own fine print.

What actually ships in the repository

Eleven folders, each holding a SKILL.md and a script: trader-scanner, trader-news, trader-entry, trader-exit, trader-timemachine, trader-risk, trader-budget, trader-broker, trader-alerts, trader-analyst and trader-pelosi. There is a tests folder with an offline unittest suite, a LICENSE, and a README. The About panel on the repository page describes the project as "11 Claude skills for day trading: scan thousands of charts, track breaking news, time every entry and exit, and replay every outcome before you risk a dollar." The sidebar records Python as 100 percent of the code.

The data is free and public by design. The README says the tools "read free public data: Yahoo Finance charts, Nasdaq Trader's symbol lists, SEC EDGAR and the House Clerk", and that there is no signup and no API key. The tools need Python 3.8 or newer and nothing to pip install. The SEC-backed skills ask you to identify yourself, with a name and email in trader/config.json or in a SEC_USER_AGENT environment variable, which is the standard courtesy for automated requests to EDGAR.

There are three ways in. Paste the repository URL into Claude and say "install skill". In Claude Code, add the repository as a plugin marketplace and install day-trader-skill from it, which namespaces the commands as /day-trader-skill:trader-scanner and so on. Or clone the repository and copy the trader-* folders into ~/.claude/skills/ or a repository's .claude/skills/, which keeps the plain /trader-scanner names. The README also notes a fourth, degraded path: paste a single SKILL.md at the top of a chat and it runs as a mode, but you lose the Python tools and keep only the method.

The author is Jake Schincariol, whose earlier trader-skill covered copying an investor's 13F filing, a stock screener, tax lots and rebalancing. This repository is the day-trading sequel, which the README calls "Trader skill v2". The split is deliberate: the first pack was about holding positions and knowing what you own, and the second is about the hours between the opening bell and the close.

The eleven commands, grouped by the question they answer

The README lists the commands in a table, which is the least interesting way to read them. Group them by the decision each one serves and the shape becomes clearer.

• Look: /trader-scanner scans thousands of charts for opening-range breakouts, flags and oversold bounces; /trader-news tracks breaking news on every stock you watch; /trader-analyst reads earnings reports and sums up what changed; /trader-pelosi reads Congress's stock-trade disclosures.

• Time: /trader-entry watches a setup bar by bar and says when to get in; /trader-exit sets the stop and the target before you enter; /trader-timemachine replays outcomes before you risk anything.

• Size and check: /trader-budget sizes a position to your budget and your maximum loss per trade; /trader-risk checks every order against your limits and stops anything over them.

• Place: /trader-broker sends approved orders to your broker, paper by default and never without your yes; /trader-alerts rings on a price, a stop or the opening bell.

Two of those deserve a longer look. The scanner works on a universe of about 5,900 US stocks pulled from Nasdaq Trader's symbol lists, and its daily mode looks for pre-market setups across the whole universe. The Time Machine is the one the README calls "the craziest one": it replays the rest of a session ten thousand times, with each five-minute candle drawn from the stock's own candles at that time of day over its recent history, and it also re-runs the rule on every past session where the setup fired.

How the pieces hand off

Everything lands in a trader/ folder in your project, and each skill reads what the last one wrote. For a single trade the chain is scanner, then news, then entry, then exit, then timemachine, then budget, then risk, then broker. While a position is open, alerts watch the stop, the target and the close. News, analyst and pelosi run on a schedule rather than in the chain. That file-based handoff is why the pack works without a server or a database, and why you can read every intermediate decision as a file after the fact rather than trusting a black box.

The repository page makes the shape obvious before you open a single file: eleven folders named after the eleven commands, a tests folder, and a commit message that reads "The Trader skill v2: 11 day-trading skills". The About panel carries the same one-sentence description as the README's first line, and the sidebar shows 637 stars, 121 forks and no releases. Stars measure attention, not effect. Three days of attention on a repository created on 20 October is evidence that people are curious, not evidence that the method works.

The TSLA example, read carefully

The README's worked example is Tesla on Monday 19 October 2026, and it is detailed enough to check the arithmetic against. At 10:15 AM Eastern, the scan covered 231 charts and 42 of them had closed above their fifteen-minute opening range. TSLA was one: an opening range of 354.47 to 362.18, with the 10:10 candle closing at 364.18, about 0.55 percent over the top. The entry timer said waiting at 10:12, because the 10:10 bar was still forming, and triggered at 10:15 on the first five-minute close above the range.

The exit planner set the plan before the entry: stop at the range midpoint, 358.32, which put 5.86 a share at risk; target at one times that risk, 370.04; flat by 3:55 PM either way. The Time Machine then replayed the rest of the day ten thousand times with seed 19, so that anyone can repeat it. Target first came up 29.4 percent of the time, stop first 25.5 percent, and 45.1 percent of replays were still open at the close. It also re-ran the same rule on the previous 59 real sessions: it fired 33 times, hit the target 13 times, the stop 15 times, and neither 5 times. The README's own verdict on that record is "close to a coin flip".

That verdict is the honest part of the example, and it is worth stating plainly what those numbers are and are not. They are the output of a simulation built from TSLA's own recent five-minute bars, replayed over one author's chosen day, on one author's chosen rule. They are not a win rate, not an expected return, and not a forecast; the 45.1 percent "still open" bucket on its own shows that the three outcomes do not describe a settled trade. One worked example, picked by the person who wrote the software, is a demonstration that the pipeline runs end to end. It is not a track record, and it cannot be turned into one by quoting it more often.

Paper is the default, and nothing fires on its own

The safety design is the strongest engineering in the repository. Paper trading is the default twice over: "paper": true is set in config.example.json, and the brokerage connector the README points at, Alpaca's MCP server, defaults to its paper account. Keys live in environment variables, never in the chat. Options, margin, short selling and leverage are off unless you turn them on.

Only one skill can lead to an order, and only after three separate checks. /trader-risk tests the exact order against the maximum loss, the daily loss limit, trades per day, reward to risk, the entry window, position size and cash, and returns PASS or BLOCK with the reason. /trader-broker then shows the ticket in plain words and waits for you to type yes for that exact order. Claude's own "Allow once / Deny" card asks a second time, provided you have set the brokerage tool to "Needs approval" under Customize, Connectors, tool permissions rather than "Always allow". A budget example from the README: on a $5,000 account with a $50 maximum loss per trade and 5.86 a share at risk, the position sizes to 8 shares, about $2,913, and a stop filled at 358.32 would lose about $46.88.

None of this is investment advice, and the repository says so itself in a section called Fine print: "Not financial advice. This software reads public data and does arithmetic. It does not know your situation and it cannot see the future. Most people who day trade lose money." That is the author's sentence, not ours, and it is the correct frame for everything above. A well-built confirmation dialog does not make a trade a good idea; it makes an accidental trade less likely.

What the pack cannot see

Four limits are documented in the README and each one changes how the output should be read. First, a stop is a plan, not a guarantee: in a fast market or across a gap, a stop order can fill well past its price, and the maximum loss shown is the loss at the stop price, not the worst loss possible. Second, the data is free and unofficial. Yahoo Finance's chart endpoint is not an official API, it can lag, change or rate-limit you, and the scanner and entry timer read one-minute or five-minute bars rather than every tick, so anything that matters should be checked against your broker before you act on it.

Third, the calendar is not neutral. FINRA removed the $25,000 pattern-day-trader minimum on 4 June 2026 under Regulatory Notice 26-10, and brokers may phase the change in until 20 October 2027, so your broker may still apply the old rule; the pack has a pdt_guard setting for exactly that case, which makes /trader-risk count your day trades. Fourth, context that looks like an edge usually is not one. Congress's trades are public rather than inside information: members must disclose trades over $1,000 within 45 days under the STOCK Act, and the Pelosi skill reads those filings when they post. Day-trading gains are short-term gains, frequent trading makes wash sales likely, and the README says plainly that none of this is tax advice.

No audited record exists, and a replay is not a prediction

This is the part to hold onto if you remember only one paragraph. Nobody has published an audited track record for this pack. The repository does not publish a return, a live account history, an independent review, or a backtest over a fixed period, and it does not say which model performed the analysis. The Time Machine is a simulation built from the stock's own last sixty days of five-minute bars, and the README is explicit that "the future can do things those 60 days never did: an earnings report, a halt, a headline." A simulated replay is not a prediction. Its odds are a reality check, not a promise, and no expected return can be derived from them without inventing the part that matters.

What the repository does publish is a test suite, and that is a different kind of evidence. The tests run offline on fixtures cut from real data: TSLA's five-minute bars for the sixty days to 19 October 2026, the 10:15 opening-range scan, that morning's Yahoo headlines, the House Clerk's 2026 index, real disclosures and real XBRL facts. They reproduce every number in the example. That tells you the arithmetic is consistent with the inputs. It tells you nothing about whether the method makes money, and the repository does not claim otherwise.

Where a model fits, and where it does not

The eleven skills are instruction files; the Python tools are ordinary scripts that run whether or not a model is in the loop. If you are wiring a model into that loop, the choice of endpoint is a separate decision from the trading method. OrcaRouter hosts models behind one OpenAI-compatible endpoint: 200-plus models, one key, provider list price passed through at 0 percent markup, with automatic failover when a provider falters. Claude Opus 5.5 sits on that catalogue at $4.00 per million input tokens and $20.00 per million output tokens, with a one-million-token context and cache reads at $0.20 per million. We host models. We do not host Anthropic's apps, Claude Code, Claude Tag, or anyone's trading skills, and nothing here changes the pack's paper-trading default or its request for your yes.

If you install it, read the SKILL.md files before you run them. They are short, they are plain text, and they are where the actual method lives, which is the part a screenshot of eleven folder names cannot show you. Then run it on paper for a while and watch what the Time Machine says about the setups you would have taken. The tool is built to make that cheap. That is the whole offer.