A wiki that keeps itself current.

SotU is a self-maintaining wiki of product insight for your research. Here's the idea behind it, the loop it runs, and what's under the hood.

The problem

Research piles up faster than anyone can read it, interview transcripts, survey exports, NPS comments, funnels. The answer to “what do our users actually need?” ends up scattered across decks and docs that are stale the moment they're written. Re-synthesizing it by hand, every quarter, doesn't scale.

The idea: a living wiki, not a one-shot report

SotU treats your research as a co-evolving research memory. Instead of wiping and rebuilding a summary each time, it maintains a small set of insight pages that compound from new evidence and corrections. The durable artifact isn't the extraction; it's the schema that does the extracting, and that schema improves over time.

This follows Andrej Karpathy's “living wiki” framing: the product's lasting value is a memory that maintains itself, not a snapshot that rots.

The loop

Ingest

Upload raw research, transcripts, exports and notes, .docx, .txt, CSV, JSON. SotU parses each interview quote, survey verbatim, NPS comment and analytics figure into a signal, with provenance back to the source file.

Curate

A curator reads each signal and makes one call: PROMOTE it into a new insight page, REINFORCE an existing page, or mark it NOISE. Corroborated themes become pages; one-off reactions and vanity metrics are held back. The result is a compact, trustworthy wiki instead of a dumping ground.

Ask

Ask the wiki in plain language, “what's blocking onboarding?”, and the answer is built from the curated pages and their evidence, not the raw pile. Good answers can be saved back as new insight pages.

Improve

This is the part that compounds. When you mark a page noise or important, the curator rewrites its own curator/SKILL.md, the rulebook it uses to judge signals. You see the before → after diff, the version climbs, and the next ingest is sharper.

Self-improvement you can measure

Where a project has held-out labels, SotU scores the curator's signal precision (with recall, F1 and junk-page counts) and tracks it across curator versions. Where it doesn't (most real uploads), it surfaces weak pages (insights resting on a single uncorroborated signal) and the climbing curator version as the proof of learning. Either way, the SKILL.md diff is the visible evidence the system is getting better.

Per project, plus a company view

Each project or study gets its own isolated workspace, its own signals, wiki, metrics and curator, so insight never bleeds between projects. The Company view rolls every project into one cross-cutting picture, with pooled metrics across the org.

Under the hood

The curator, answerer and linter are each a small, inspectable skill file, the curator's is the one that rewrites itself.

Try it on your own research.

Sign up, create a project, drop in a file, and watch the wiki build itself.