The Trading Assistant Whose Best Feature Is Telling You It's a Coin Flip

There is a genre of trading software that promises to make you money. Jake Schincariol’s day-trader-skill belongs to a much rarer genre: software that promises to tell you the odds, and then gets out of the way. The repo — eleven Claude skills covering everything from chart scanning to congressional trade disclosures — pulled 831 stars and 178 forks within weeks of landing on SkillsLLM’s featured rail in late October 2026, and the reason it travels is not that it trades for you. It’s that it refuses to.

The project is a sequel. Schincariol’s earlier trader-skill handled the investor’s side of the ledger: copying 13F filings, screening, tax lots, rebalancing. This repo is its day-trading counterpart, and it inherits the same design philosophy — plain Python, free public data, no signup, no API key, nothing to pip install. The tools read Yahoo Finance charts, Nasdaq Trader’s symbol lists, SEC EDGAR, and the House Clerk’s disclosure feed. That’s the entire supply chain. In a market where the competing platforms charge $127 a month for AI scanning, the zero-dollar supply chain is itself a statement.

A pipeline, not a chatbot

The architecture is the unglamorous part, and it’s where most of the value lives. Each of the eleven skills is a folder with a SKILL.md and a small Python tool, invoked as a slash command — /trader-scanner, /trader-news, /trader-entry, and so on. The trick is that everything lands in a shared trader/ folder in your project, and each skill reads what the last one wrote. Scanner output feeds the news monitor; the news monitor feeds the entry timer; the entry timer feeds the exit planner; the exit plan feeds the simulator; the simulation feeds the position sizer; the sized order feeds a risk checker; and only then does anything approach a broker.

This is a deliberately boring design, and it’s smarter than it looks. The fashionable approach to agent-based trading is one omniscient model that scans, decides, and executes in a single conversational breath. Schincariol instead built a state machine with human-legible checkpoints. Every intermediate artifact — the scan, the plan, the simulated odds, the sized order — exists as a file you can read, question, or throw away. The model orchestrates; it does not improvise. When something goes wrong mid-pipeline, and in trading something always goes wrong mid-pipeline, you can see exactly which stage produced the bad number.

The pipeline also enforces a discipline that most retail traders skip: the exit plan is written before the entry, the position is sized to a fixed maximum loss before the order exists, and a risk checker issues a binary PASS or BLOCK with a stated reason. The example trade in the README — Tesla, October 19, 2026, real data — walks the whole chain: an opening-range breakout at 364.18, a stop at the range midpoint, a target at one times the risk, eight shares on a $5,000 account with a $50 max loss, a $46.88 worst case. The arithmetic is trivial. That’s the point. Most trading mistakes are not exotic; they are arithmetic nobody did.

The Time Machine

The skill that gives the repo its identity is /trader-timemachine, and it deserves a careful description because it is either the most honest or the most seductive tool in the pack, depending on how you read it.

Before you take a trade, the Time Machine replays the rest of the trading session 10,000 times. Each simulated 5-minute candle is drawn from the stock’s own candles at that time of day over the previous 59 sessions — so the simulation inherits TSLA’s actual intraday temperament, not some generic random walk. Then, separately, it re-runs the same entry rule across those 59 real days and reports how the setup actually performed historically. The run is seeded — seed 19 in the README’s example — so anyone can reproduce the exact numbers.

On the Tesla trade, the verdict was: target hit first 29.4% of the time, stop hit first 25.5%, still open at the close 45.1%. The historical replay found the setup fired 33 times in 59 days, hitting the target 13 times and the stop 15. The tool’s own summary: close to a coin flip.

This is where the project earns its keep, because a coin flip is exactly what most breakout setups are, and almost no trading software will tell you so. The commercial ecosystem — TrendSpider, Trade Ideas, DayTrade AI — is built to surface setups, draw the analysis, hand you entry and exit levels. The economic incentive runs toward confidence. A tool whose headline output is “this is roughly a coin flip, and 45% of the time you’ll still be holding it at the close” is working against its own adoption funnel. The README’s framing is unusually direct: a coin flip is not a reason to skip the trade or take it; it is a reason to know what you are buying before you buy it.

Technically, the Time Machine is a bootstrap Monte Carlo with a time-of-day conditioning — resampling the stock’s own recent bars rather than fitting a parametric model. It is not sophisticated quant methodology, and it doesn’t pretend to be. The README’s fine print says it plainly: it is a simulation, not a prediction, built from the last 60 days of bars, and the future can do things those 60 days never did — an earnings report, a halt, a headline. What it displaces is not a research desk; it’s the trader’s imagination, which is a notoriously biased simulator.

The consent architecture

The other design bet is layered human approval. Nothing in the pack can place an order on its own. Only the broker skill can lead to an order, and only after the risk checker has passed that exact order and the user has typed yes in the chat. Then Claude’s own permission card — “Allow once / Deny” — asks again. Paper trading is the default; the brokerage connector is Alpaca’s MCP server pointed at its paper account; keys live in environment variables, never in the chat. The README even instructs users to set the brokerage tool to “Needs approval” and never “Always allow.”

Three separate gates — risk check, typed yes, permission card — for a retail day-trading tool is almost comically cautious. It is also the correct read of the moment. The StockBrokers.com guide to AI trading platforms distinguishes between AI-assisted research apps and true automated bots, and its standing advice is to treat any AI trading tool as an aid to judgment, not a replacement for it. Day-trader-skill hard-codes that advice into the architecture. An agent that could silently place orders on unofficial data from a rate-limited Yahoo endpoint would be a liability generator; this one is built like a checklist with a language model attached.

The boring stack, tested offline

Worth pausing on: the tools are all standard-library Python, 3.8 or newer, with zero dependencies. In 2026, a trading toolkit with no pip installs and no vendor SDKs is nearly a contrarian position. It also means the test suite runs offline, on fixtures cut from real data — TSLA’s 5-minute bars for the 60 days to October 19, the actual 10:15 AM scan, that morning’s headlines, the House Clerk’s 2026 index — and the tests reproduce every number in the README’s worked example. Reproducibility from seeded simulations plus fixture-driven tests is more scientific hygiene than most trading software, free or paid, can claim.

The Pelosi tracker, meanwhile, is the pack’s viral-friendly garnish: it reads congressional trade disclosures the minute they post, per the STOCK Act’s 45-day rule. The README is careful to note these are public filings, not inside information — a distinction the internet has never once respected.

What it doesn’t do

The honest limits are mostly stated by the project itself, which is refreshing. The data is free and unofficial: Yahoo’s chart endpoint is not a contracted API, and it can lag, change, or rate-limit you. The scanner and entry timer read 1- or 5-minute bars, not every tick, and the README tells you to verify anything that matters against your broker. A stop is a plan, not a guarantee — fast markets fill stops well past their price. The simulation window is 60 days, which in a regime change is a history of a world that no longer exists.

The deeper criticism is structural, and the project doesn’t dodge it: this is software that reads public data and does arithmetic. It has no alpha. The setups it scans for — opening-range breakouts, flags, oversold bounces — are the most heavily traded patterns in retail markets, and a coin-flip verdict on a TSLA breakout is consistent with everything serious research says about simple technical rules. There’s a live debate, visible in the algotrading community, about AI making it easier to build backtests that look real but aren’t — and a 10,000-replay Monte Carlo is exactly the kind of artifact that can manufacture false precision. The mitigations here are real (seeding, the historical replay alongside the simulation, the coin-flip framing), but a user who wants to be fooled can read “29.4% target, 25.5% stop” as an edge rather than a warning. The tool can’t fix the reader.

There’s also a timing wrinkle the project handles with unusual care: FINRA removed the $25,000 pattern-day-trader minimum in June 2026, but brokers may phase the change in until October 2027, so the old rule may still apply. The risk checker can count your day trades and enforce the limit either way — a small detail that suggests the author expects small accounts, the exact population most likely to lose money day trading, and the exact population the README’s “most people who day trade lose money” warning is aimed at.

Where this sits

The commercial landscape around it is consolidating toward subscription platforms: TrendSpider sells an all-in-one research suite with an AI workbot; Trade Ideas charges $127 a month for its Holly AI signals; DayTrade AI sells chart-reading analysis at $8 to $20 a month, with the AI explaining numbers that “our analysis engine” produces. Day-trader-skill occupies the opposite corner: free, MIT-licensed, inspectable, running on your machine against public endpoints, with the model as orchestrator rather than oracle. It will not compete on features — no real-time tick data, no institutional-grade charting, no predictive ML lab. It competes on trust and transparency, which in this product category are nearly unclaimed territory.

The open question is what the skills ecosystem does with it. The pattern here — a pipeline of small, file-passing skills with human checkpoints — generalizes well beyond trading, and the repo is as much a demonstration of that architecture as it is a trading tool. Schincariol’s own site frames it as part of a broader playbook library, with a companion resource page and a video walking through the backtesting flow. If the pattern spreads, the interesting legacy won’t be the trading desk. It’ll be the proof that an agent system can be most useful precisely where it’s built to say no.