Join the Latticework AI Bootcamp and progress at your own pace. Participation is open to members and paid subscribers.
A note before we begin: This is the tenth lesson in a 16-lesson experiment. I am doing every lesson alongside you, on the same tools, with the same constraints. Some lessons will land cleanly. Some will lead to dead ends and need rework. We’ll figure out what works, together.
If you are catching up, here’s what came before today’s lesson:
Intro: Build Your Own Investment Idea Engine
Lesson 1: How LLMs Work, and How to Defend Against Hallucinations
Lesson 2: Prompt Patterns That Outperform Casual Prompting
Lesson 3: Tools, Agents, and Structured Output
Lesson 4: SEC EDGAR, the Primary Source
Lesson 5: FRED Macro and Sector Data, the Free Read
Lesson 6: FMP API Key and the First Checked Data Pull
Lesson 7: Other Data Sources, and Idea Engine Formats
Lesson 8: Spaces, Projects, and Connectors
Lesson 9: Memory, Sub-Agents, and Parallel Research
Why This Matters for Investors
In the last lesson we ran five parallel sub-agents on five companies, with the same schema and the same source rule for every row. Today we widen the same pattern from five names to twenty. The shape of the work does not change. What changes is that we are no longer treating wide research as a thinking exercise on a small set of companies we already know well. We are using it as a triage layer across a small universe, the kind of universe we would not have time to read through one company at a time.
A classic quantitative screen narrows the field on numbers. Twenty companies above a ROIC threshold, twenty inside a market-cap band, twenty with insider ownership over a line. That is useful, but it is also coarse. It leaves out everything we actually have to read to know whether a name is worth an hour: the recurring concerns inside the last filing, the customer concentration buried in a risk factor, the management quality signals that only show up across two or three years of letters, the ownership patterns that change how an activist or a founder is likely to behave.
“Wide research” adds that qualitative layer at scale, but only if we keep it source-grounded. The reason most members get into trouble here is exactly the same reason most members get into trouble with any AI-assisted research: a smooth comparison table is not evidence. A smooth comparison table is the assistant doing what it is best at (writing fluently across many rows), in a domain where fluency is the failure mode we have been defending against since Lesson 1. So our job today is to set up the run so that the table cannot lie to us without us catching it.
There is a quieter benefit as well. The discipline of writing one schema, one source rule, and one universe definition before any names are pulled is good for our own thinking. It forces us to say out loud what we are actually looking for, and what we are willing to call “missing.” Twenty rows under a strict schema beats two hundred rows we cannot audit.
Let’s launch into today’s lesson.
