Ian here, ferreting the internet like a truffle pig looking for nomnoms. 🐷 This is an example of my own personal automated trading strategy (binary coded), adjusted for an equity curve + guard rails specifically designed for passing propfirm evals. Tests say it should be a 100% pass rate, but we'll see about that! Just a fun side-experiment I wanted to try. BUT: The base strat currently runs on my live account, built from my own trading experience and what I personally see on the charts. It was rough, going back and forth with Claude co-work, but it definitely did the heavy lifting of building Python scripts for testing, parsing 16 years worth of historical data, and coding/re-coding it to match my typical strategy. Took a long time and many dollars. 🤑 My objective is to move towards a more agentic style, where it can watch the trades that are taken autonomously, understanding what might need to be changed- Then implementing those changes for more testing. Plus, self-learning experimentation to improve upon the base strategy. Not new to LLM use, but never made a single agent before. 🙃 FYI, the PnL looks astronomical, but it's designed to be a consistent grinder. This just shows how consistent it actually is starting with only $2k! On a $50k account, the drawdown is only 2.64% (as built for my challenge)...

I have just completed Days 1–16 of the Zero One Systems curriculum. My contribution has been applying the prompts provided each day to my own work, challenging the assumptions where they did not fit, and following the process far enough to see what emerged. For my use case, a personal agent is not simply a chatbot that remembers you. It is an operator-facing manager backed by explicit context, bounded authority, specialist systems, and evidence. That is the claim this post is trying to earn. I began with a personal dashboard and a simple question: what would an AI need to know about me to become genuinely useful? The curriculum works through personality, values, goals, risk tolerance, and decision-making. The most useful—and uncomfortable—exercise was a documentary-style interview covering my background, career change, failures, family, money, and what actually drives me. That became soul.md: a private canonical file describing how I think and operate. I then distilled it into soul.runtime.md, a smaller set of executable principles for practical agent use. Two of those principles have already changed the architecture: - Autonomy requires bounded authority, which exposed the weakness in my original agent design. - Correctness governs speed, which is why completion evidence now belongs in an append-only - Decision and Evidence Ledger rather than being reduced to a status flag. One lesson mattered more than the rest: More context is not automatically better. Stable identity, current project state, private history, operational knowledge, and evidence are different classes of information. They update at different rates and should only be exposed to agents that genuinely need them. The biggest change came when I reviewed which agent to build first. My initial choice was a Founder Intelligence Scout—monitoring AI tools, GitHub repositories, contracting opportunities, and founder tactics. After two separate research workflows, the problem became obvious. I already had specialised systems doing adjacent work:

I have a Personal Claude.md file now.