Most analysts are using AI backwards. They chase the newest model. Which one scores higher on the benchmark, which one just shipped, which one their friend swears by. Wrong question. Here is why. Everyone has the same model. You, me, the analyst across the street, the algo desk down the hall. We are all renting the exact same frontier model. So if the model were the edge, every research note in the market would read the same. They do not. Your edge is not the model. Your edge is the context you feed it.
And most people feed it through ten disconnected chat tabs. One tab for the 10-K. Another for the earnings call. A third for the thesis memo. Nothing persists. Every new session starts cold. The model is stateless. Open a fresh chat and it is a complete beginner. It does not know your coverage. It does not know what you concluded last quarter. The real cost is the scavenger hunt. You spend twenty minutes re-finding the note, the source, the conclusion you already reached. That is the failure mode. Scattered context, zero memory, no audit trail.
Here is the fix. One research workspace, built on four layers. Each layer depends on the one before it. Build it bottom-up.
Layer one. Context. This is the foundation. The workspace knows your coverage. You open a session and ask, what do we track, and what did we conclude last quarter. And it answers from your own notes, not generic web text. Where does that live? In a memory file. A CLAUDE.md, a coverage file. Your names, your thesis, your prior conclusions, written down once so every session starts as an analyst instead of a beginner. No memory file, no edge. Start here.
Layer two. Connections. This is what the workspace can actually read. Not copy-paste. Direct access. Your filing folder. Your transcript folder. Your research drive. A monitoring feed. You point it at the 10-K folder and it reads the filing directly. And here is the rule that matters. Scope every connection read-only. Assume that if your agent can read something or do something, it will do it. Not what you told it to do. What it can reach. So market data, filings, transcripts: read-only. Execution: not connected at all. You hand it the keys you choose, and nothing else.
Layer three. Capabilities. These are your repeatable moves, saved as skills. The way you read a 10-K. The way you summarize an earnings call. A fact-versus-inference pass that separates what management said from what you are assuming. You build a skill the way you teach a kid to ride a bike. First you supervise. You watch it run, you correct it. Then, once it is reliable, you grant autonomy. Do not skip the supervision step. A skill you have not corrected is a skill you cannot trust.
Layer four. Cadence. This is work that runs on a schedule, so monitoring happens without you asking. A coverage monitor that checks for new filings every morning. Scoped, read-only, running in the background. But it only works if the three layers underneath it exist. Cadence on top of bad context just automates the mistake faster.
Let me show you what one layer actually produces. Take the 10-K read skill. You run it on a filing, and it gives you a structured note. Prior period versus current. The deltas: revenue drivers, margins, guidance, risk factor changes, shifts in management language. Each point cites the source, the exact filing section, and the date. And it ends with a "what to verify next" list. That is the analyst standard. Source, section, date, and an open-questions list.
Now notice what this is not doing. It is not telling you to buy. It is not telling you to sell. It read the filing faster than you could and laid out the changes. The judgment is still yours.
Let me be direct about where this breaks, because it does break. The model can be confidently wrong. It will state something with total certainty that is simply not in the filing. That is why every output traces to a source. If a claim has no source, section, and date, you do not trust it. You verify it.
And the connection layer is where people get burned. The common story is an agent that picked up a task on its own and acted. It sent emails nobody approved. The lesson is not "write a better instruction." Telling your agent "never send emails" is not the same as removing the send key. Instructions are not capabilities. If the key is on the key ring, the door can open. So you do not forbid it. You remove the key. Market data and filings, read-only. Execution, out of the loop entirely. You control the permission layer, and the agent controls nothing you did not hand it.
So here is how you use this safely. Keep every market connection read-only. Keep execution disconnected. Put a source, a section, and a date on every claim. And keep the judgment with you. Because you can outsource the work. You cannot outsource the understanding. The output is a starting point, never a substitute.
One last thing, because it is the whole point. A research workspace does not make a weak thesis right. It will not save bad analysis. What it does is make a disciplined process faster and easier to audit. That is the edge. Not the model. The context, and the discipline around it.
This is educational only. It is a workflow example, not a recommendation to buy, sell, or hold anything. Markets involve risk. Verify your data and do your own work. If you want the analyst AI research-stack checklist, comment STACK and I will send it over.