What we believe
I keep watching the same scene. A PM I respect opens her laptop, scans seven dashboards, three Slack threads, and a half-finished doc, then closes everything and goes for a walk. The walk is where the work happens. Not because the tools failed. They gave her too much. None of them could tell her which piece mattered for the decision in front of her.
I've watched this scene thousands of times. The tools have gotten extraordinary at producing context and useless at exercising judgment over it.
Aristotle had a word for what she was doing on that walk. Phronesis. The trained eye for what this situation calls for... when no rule fits and no past case matches. It used to come through apprenticeship: long hours next to someone who already had it. That mode mostly doesn't exist anymore, and the tools we built for the new conditions made the production problem easier and the judgment problem harder.
Here's the part that changes everything in the next few years. The same gap that traps the operator traps the agent the operator might delegate to. Today's agents act from thin context: a prompt, some retrieved chunks, a half-readable history. They make plausible-looking calls that anyone with real judgment and a grasp of the specific situation would catch. The workspace that solves the human's judgment problem is the same workspace that gives an agent something worth working with.
That's what we're building.
What we're building
Craft Phronesis is one product: a workspace where judgment compounds.
One engine (mesh, ambient capture, Layer 1 / Layer 2, skills as rituals, compile and retrieve loops) applied to whatever domain the operator works in. PM. Personal life. Research. Engineering. Anything that benefits from sustained judgment over time. Not an umbrella with launch verticals under it. One product, with a door for each domain. The architecture is the architecture. The entry point is the entry point.
What the engine scaffolds is the part that lasts. Not the expert's playbook, the part frontier models are already learning to extract from interaction logs. The situated judgment a skilled operator brings to a conversation that will not repeat: which thread to pull, which thirty seconds carried the load. It compounds that judgment, session over session, into a workspace the operator owns.
The pattern is not ours to invent. Live AI with context has already proven itself in sales, in the room, on the call. We are porting it to the discovery interviews, stakeholder 1:1s, and customer conversations operators actually run, and producing something structurally different at the end: a compounding personal workspace, not a closed-loop vault someone else keeps.
This is judgment, not retrieval. The workspace holds your caught cases, the raw and compiled record of what you've encountered. The codex holds your aim: the taste you've built, the principles you've validated, the ends you've come to trust. Retrieval brings back what you said. Judgment knows which of it matters here, and toward what. Aristotle's word for skill without the right aim was cleverness. The codex is what keeps this from being merely clever.
An operator picks their domain, walks through the door, and the system bootstraps a workspace populated with first-party content from people who already have phronesis in it. Same engine behind every workspace. Each gets richer as the operator lives in it. The engine sharpens across all of them as it's used in any one. And every workspace becomes something an agent can act from. That's the flywheel.
What it feels like
Your thinking compounds, and keeps working after the moment that needed it.
You make a hard call on Tuesday. By Thursday you'd have forgotten the reasoning. The workspace caught it. When the next call rhymes with the last one, your prior thinking is there. So is the codex.
It builds from small moments. The "I saved that" beat when the workspace catches a decision you'd otherwise lose. The compile that turns scattered notes into something you recognize as yours but didn't have to write. The agent that pauses you before a wrong move because it knows what you concluded six weeks ago even though you can't recall it right now.
Eventually the agent acts on your behalf, from a workspace rich enough that it has something to act from. Your taste, encoded. Your prior thinking, available. The codex, woven in. Not a smarter agent. The same agents everyone has, given something worth working with.
You don't experience features. You experience your own judgment with full memory of your own work, at the moment you need it.
Who we're for
Operators reaching for thinking-tools and not finding them. People who tried five second-brain setups and stuck with none. People who feel the gap between volume and judgment and don't have a name for it.
We start with technical PMs already living in AI-native workflows. They're the audience closest to the problem. They'll feel what the architecture is doing the first time they use it, and they'll bring the rest.
What we won't do
We won't add features to compensate for an unsharp core. When we reach for "one more thing," the core isn't exceptional yet. We sharpen instead.
We won't expand to a new domain until we've nailed value for one.
We won't open the ecosystem until we've earned the audience that populates it.
We won't lock you in. The workspace is yours, in a format you can read, on a disk you control. Whatever we build, you can walk away with, and run any model against.
We won't join a category. We're naming one.
We won't dress rule-following as judgment. False precision from subjective inputs is not judgment. Situated reasoning is. We design for the second one.
What we're aiming at
A category named for what every serious operator wants and no current tool delivers: a workspace that cultivates judgment in their domain, for them, over time.
There is no moat for us, and that is the design. The workspace they build is theirs. The codex they shape is theirs. Whatever frontier model arrives next is just better fuel for the workspace they've already built. We don't have to beat the frontier. We have to be the layer where their taste accumulates, regardless of which model runs behind it. We help operators own their own data so completely that they stay free to run any tool against it, including ours, or the next one.
If we get this right, we don't compete on volume. We compete on the gap between volume and judgment, the gap everyone who's tried AI or a second brain has felt and couldn't name.
That's Phronesis.