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 eleventh 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
Lesson 10: Wide Research for Screening at Scale
Why This Matters for Investors
In the last lesson we ran a twenty-name wide-research triage pass under one strict schema; and the schema, not the universe, was the thing that made the table honest. Today we take that lesson one layer deeper. A prompt is not a clever sentence. A prompt is a small operating procedure that has to behave the same way on the fifth ticker as it did on the first, on a sector we know cold and on one we do not, on a name with a clean filing history and one with a recent restatement. If we do not test it deliberately, we will keep noticing failures only when they show up inside an investment decision.
There is a second reason to slow down here. Most non-technical members get into trouble in the same way: they write a prompt that produces one good output on a ticker they already understand, conclude that the prompt works, and then are surprised when the same prompt produces a polished but unsourced write-up on a name they do not know. That second output is the dangerous one. It will be fluent, it will be plausible, and it will smuggle in assumptions about discount rates, terminal margins, share counts, or filing dates that the model invented because the prompt did not forbid it. Prompt iteration is how we close that gap.
The other half of today is the structured idea write-up itself. Research that does not land in a comparable shape is hard to compound. If every company note has a different structure, we cannot scan across them, cannot revisit our reasoning later, and cannot tell whether our process is improving. A five-bullet house-style write-up, generated from a tested prompt, gives us that comparability without flattening the thinking. And once the write-up exists, we have earned the right to run a simple reverse DCF on one name we know well, where the discipline is to separate market-implied assumptions from our own and to refuse to let the model invent inputs.
By the end of today we will have a small prompt-evals sheet with five rows, one documented failure pattern, one targeted prompt fix, and one reverse-DCF output whose inputs we have personally checked.
Let’s launch into today’s lesson.
