THE GROUND ZERO KIT

Your AI isn't stupid. It's unsupervised.

Four files. Free. Nothing to buy on this page.

The problem

You have probably had this exact experience.

You ask an AI to build something. It builds it. It looks right. The code runs, the explanation is fluent, the numbers in the summary are specific and confident. You ship it.

Then two weeks later you find out one of those specific confident numbers was never computed. It was generated. It sat inside a well-written sentence and you believed the sentence.

That is the failure mode. Not "the AI can't code." It codes fine. The failure is that a language model will produce a plausible answer with exactly the same tone of voice whether or not the answer is true, and it will agree with whatever you seem to want, and tomorrow it will happily re-propose the idea you already proved was dead last Tuesday.

I build trading software. That is a domain where being confidently wrong costs actual money in a way you can measure. So I spent six weeks building something that isn't a strategy and isn't a model. It's a set of files that sit between me and the AI and force both of us to prove things.

Snapshot from my repo on 26 July 2026: 243 commits in 44 days, 890 Python files, 185,766 lines, 315 notes in the knowledge vault, 153 memory files, peak day 24 commits. Most of what's in there is stuff that didn't work, written down so I stop rediscovering it.

Two things that came out of it, so you know what this process is actually for:

1. A chart panel told me a strategy arm was running 24 wins and 4 losses over 14 days. I rebuilt it in Python to check. The rebuild matched the chart exactly, 28 of 28 signals, same prices, same minutes. Then I resolved the trades against real fills instead of the chart's own logic. 13 wins, 15 losses. Negative. The chart had been scoring each trade against price action from before the trade existed. The 86% was a bug. I nearly traded it.

2. The original goal of the whole project was a 70% win rate. I hit it. 72.8% in-sample, 71.2% out-of-sample. Then I checked what it was actually doing: risking three units to make one. I deleted it, because it broke a risk rule I'd written down before I ran the test. Separately, a random-entry control on the same data produced a higher win rate than my real strategy, 49.6% against 47.9%. My entries were worse than a coin flip. The edge, such as it is, was somewhere else entirely.

None of that gets caught by being smart in the moment. It gets caught by having written the rules down first, in a file the AI reads every single session, before it has any idea what answer you're hoping for.

What's in the Kit

Four markdown files, in one PDF.

Number 1

CLAUDE.md starter template The instruction file. It loads automatically at the start of every session, so the rules exist before the work does. The template has the sections that actually earn their place: hard rules that can't be argued with, how to break a task into pieces with a pass/fail test on each one, where to spend the extra effort, and a final checklist the model runs before it's allowed to answer you. Fill in your domain, delete what doesn't apply.

Number 2

Memory system starter How to give an AI a memory that isn't just a longer chat log. One fact per file. Every file has a name, a description, and a type. An index file on top that gets read first. The important part is the graveyard: dead ideas get a marker so they never get proposed to you again. Mine has a stop sign on things like "the ES divergence gate" and "tick absorption," and I don't have to re-explain why those are dead. Ever.

Number 3

Six questions that catch your AI lying to you The short one. Six questions you can paste into any chat, right now, with no setup. They're each aimed at a specific way a model produces a wrong answer that reads like a right one. Two examples so you know the flavor:

"Point to where in this session you derived that number. If you can't, delete it and re-derive it."

"Did you actually open that file this session, or are you citing it from memory?"

Every one of the six exists because that exact failure happened to me and cost me time.

Number 4

My real CLAUDE.md, scrubbed The one running against a live trading system, with the account numbers and API keys taken out and nothing else changed. It is not tidy. It has a section in all caps at the top telling the AI not to judge performance by dollar P&L, because it did, and it was wrong, and it needed to be shouted at. That section is there because of a real mistake. You're getting the working version, not the demo version.

Who is this for

You use Claude, Cursor, Codex, or similar to build something more than a weekend toy, and it's starting to get away from you.

You've caught your AI making something up at least once and it giving you false hopes and results.

You want the process, not the product. You'd rather know the thing you are trying to build is possible before even starting.

You're fine with being told an idea you like is dead and the only thing you lost was time not real capital.

Who this is not for

You want a trading bot. There isn't one in here. There is no strategy, no signal, no entry, no indicator. The kit is about how to make an AI build software honestly. That's it.

You want a prompt that makes the AI 10x smarter. Not a thing. This is bookkeeping and discipline, and it's slightly boring.

You want returns, a track record, or someone to tell you what to buy. Wrong channel entirely. See the disclaimer at the bottom.

You want a GitHub repo to star. My repos aren't public and aren't going to be.