Honest limits

The whole product is one claim: it does not lie. So here is everything it can’t do.

It only reads two assistants

Your evidence is built from Claude Code and Codex sessions, and nothing else. If most of your work happens somewhere else, there is little for stratless to read and it will have little to say about you. More readers arrive one at a time, the rest of the CLIs next and the GUIs later.

Delivery is per-tool work too, and both readers have their doors: the sitting installs each pair’s skills through its own tool’s native skill directory. A future reader ships with its own measured door the same way. See Assistants for where yours stands on both halves.

It can only read what survived

stratless reads all of your history, but only what still exists. Claude Code deletes transcripts after 30 days, so anything that aged out before you ran init is gone for good, and the evidence is built from what remains. That is why init is the one step you cannot do later. Codex keeps its rollouts indefinitely, so nothing there is on a clock.

The evidence is derived, and a model can be wrong

It is not a transcript. It is written by an LLM from your real history, told to say less rather than invent, with every number counted by code so none of it can be fabricated. But it is still a model of a person, reasoned from evidence. It can be wrong, and it can echo itself: if it decides something about you and the assistant then leans into it, your reactions can quietly confirm a read that was off. Open ~/.stratless/HUMAN.<assistant>.md and read it; if it doesn’t sound like you, that is the most useful thing you can tell us.

It fits the model, it doesn’t fix it

stratless doesn’t make your assistant smarter. It gives the assistant a model of you, so it stops explaining at the wrong altitude. That is the whole job, and it is smaller than “AI that gets you.”

And it’s n = 1

stratless was built by one person, from one person’s transcripts. The problem is real for him with total certainty, and for everyone else with none. If you run it and the skills look nothing like you, open an issue. That is the most useful thing you could possibly tell us.