Why Customization Is the Real Point of Tesseract Stock Agent
Most AI tools hand you a fixed thing. You get the interface the company built, the steps it chose, the output it decided you wanted, and your only real freedom is which question to type into the box. That works right up until the model provider ships the same feature for free, and the tool you were paying for quietly stops having a reason to exist. The product that survives that is the one you can take apart and rebuild. Tesseract Stock Agent was designed to be taken apart.
Underneath the polish, the system is deliberately simple. It is a chain of prompts run in sequence, each one fed by a set of curated resources, ending in a structured report. None of that is locked. Because the architecture is exposed rather than hidden, every layer of it is yours to change, and that is not a side feature. It is the whole point. Here is what customization actually means in practice.
What you can change
The resources. Each step in the chain draws on its own specialized resource files: valuation methods, sector logic, and the checklists of what tends to go wrong. Swap them. Drop in your own valuation framework, your own sector notes, your own definition of a red flag, and the analysis starts reasoning from your standards instead of the defaults.
The number of prompts. The standard chain is ten steps because ten is what a full thesis usually needs. It is not sacred. Cut it to five for a fast first look, extend it to fifteen for something forensic, or delete the steps you never use. The length should match the job, not the other way around.
The paths. The chain does not have to be a straight line. Run the optional steps only when a situation calls for them, fork the flow for a specific scenario, or build a branch that triggers a deeper dive when something looks wrong. You decide where the workflow bends.
The final report. The same analysis can end as a one-page memo a beginner can read or a ten-page breakdown built for a committee. Tone, length, and the look of the exported PDF are all adjustable. The research underneath stays constant, and only the surface changes to fit who is reading it.
The entire purpose. This is the part most people miss. The engine does not actually care that the subject is a stock. It is a structured, resource-backed prompt chain, and that pattern works on any complex analytical task. Point it at financial modeling, at deal structuring, at a due-diligence process that has nothing to do with public equities. Replace the prompts and the resources and you have a different system running on the same proven skeleton.
The data layer. The chain runs well on its own, and it runs better when you feed it live numbers. Wire in market or filings APIs where accuracy matters most and leave them out where it does not. That is a choice you make per use case, not a setting someone else locked for you.
The way it runs. Drive the chain by hand, let the Relay app run it inside your chat, or wire it into an automation workflow so it executes on its own. The logic is identical in all three. Only the effort on your side changes.
A warning that belongs here
Customization is not for everyone, and pretending otherwise would be dishonest. The defaults exist because they are good, and most users should leave them alone. Reshaping the chain, rewriting resources, and building branches is powerful in the hands of someone who knows what a good research process looks like, and dangerous in the hands of someone who does not. Freedom without competence is just a faster route to a confident, wrong answer. This layer of the system is built for experienced investors and people who work in finance for a living. If that is not you yet, the standard workflow will serve you better than anything you would build by guessing.
Why this is the moat
Most AI products cannot be customized in any way that matters, because there is nothing underneath the interface to customize. Strip the branding off and you find one prompt and a billing system. When the underlying model improves, the wrapper has nothing of its own to fall back on, which is why so many of these tools have the lifespan of a season.
A system you can rebuild is a different kind of thing. The prompts, the sequence, and the outputs are yours, down to the subject the system analyzes, which means it bends toward your process instead of forcing you into someone else's. That is the line between renting a tool and owning a system, and it is the reason the interface was never the real product here. The architecture was. Everything you are allowed to change is the proof, and it is all up to you.
