Most analysts are arguing about the wrong thing. Which model is best. Which one tops the leaderboard this quarter. Which version dropped last week. That is the wrong argument, and it quietly costs you hours.
Here is the failure mode. You open a fresh chat. You paste an earnings transcript and ask for a read. The model knows nothing about you. It does not know your coverage names. It does not know your memo format. It does not know what you consider a red flag. So it guesses. You get a confident, generic answer with the wrong flavor, and you spend twenty minutes re-explaining your own process. Same model your colleague is using. They get a clean comp table. You get noise. The model is not the difference. The wrapper around it is.
Let me define the wrapper, because that is the whole game. The wrapper is everything you put around the model. The context, the workflow, the source rules, the examples, the success criteria. Same model, very different output, depending on what you built around it. The real skill is not writing one perfect prompt. It is building the right environment. Call it context engineering.
Here is how I would build it as an analyst. Five layers. One at a time.
Layer one. Context. Write your process down. Your coverage universe. Your preferred memo format. Your valuation style. Your risk checklist. Your kill criteria, the conditions that make you drop a name. The model cannot give you process-specific output if it does not know your process. So you stop making it guess. You hand it the process in a file.
Layer two. A skill. Take one repeatable read and turn it into a written workflow. The filing comparison you run every quarter, for example. Write down when to run it. What sources to pull. What the output should look like. What caveats to flag. Now it runs the same way every time, whether it is you running it or someone on your team. Consistency, on purpose.
Layer three. A knowledge base. Not a flat pile of files you search through. A folder of source notes plus a schema file, so the agent builds a living, connected research base for a name. Earnings notes link to filings. Filings link to your memos. The agent synthesizes across them instead of grepping one document at a time.
Layer four. An autonomous loop. This is the one people get wrong. Stop saying do this one step. Start saying keep going until this condition is true. Define the outcome. Define the objective check. Let the agent iterate against it. Run until the comp table reconciles, then come back and review. But heads up. Only wrap something in a loop after the single-shot version is already reliable. A loop optimizing the wrong objective just produces precise garbage, faster.
Layer five. Human judgment. You own the assumptions. The growth call. The margin call. The multiple. And the most important question, what would make this wrong. The stack never outputs a buy or a sell. It builds, it formats, it cites. You decide.
Let me show you what the output looks like, so this is concrete. A filing comparison. Prior period versus current period, side by side. The deltas in the revenue drivers. The change in margins. The shift in guidance language. Changes in the risk-factor section. Capital allocation. And here is the part that matters. Every single row cites the source, the date, and the exact filing section it came from. If it cannot be traced, it does not go in the memo. Notice what this does not do. It does not tell you to buy or sell. It structures the work. You make the call.
Now the part nobody wants to hear. Where this breaks. No context gives you confident, generic, wrong-flavored output. A stale or wrong source is worse, because the wrapper will launder a bad input into a clean-looking memo. It looks polished. It is wrong. That is the dangerous failure mode, because the polish hides the error. And there is a quieter cost. The more context you build, the more switching cost you create. That lock-in is real. It is useful, but own it deliberately. Do not pretend it is free.
So here is how to use it safely. Every output follows the same rules. Compare against the prior period. Extract the deltas. Cite source, date, and filing section on every line. Classify each change as positive, negative, or unclear. And end with what to verify next. The AI owns construction and speed. You own the assumptions and the decision.
Step back and look at where this is going. The model is becoming a commodity. Everyone has access to roughly the same frontier model. Your context is not a commodity. Your coverage notes, your memo format, your kill criteria, your source discipline, that is yours, and that is what produces a different output from the same model. A better prompt is luck. A better wrapper is an edge.
This is educational only. It is a workflow example, not a recommendation to buy, sell, or hold anything. The AI structures your research. It does not pick your stocks, and it does not change who is responsible for the call. That is still you.
If you want the version of this I actually use, comment WRAPPER and I will send you the one-page Research Operating System checklist. The five layers, on a single page. Context, skill, source check, loop, and the human judgment that owns the decision.