Mari.guru
One Pane
Of Glass.

The bottleneck moved.

AI made writing documentation effectively free, which pushed the hard part downstream. The job was no longer generating docs; it was proving any of it was still true. The mandate: turn Mari into a transparency layer that holds every machine-written claim accountable, sourced, approved, fresh, and compliant, with a visual identity strong enough to compete on distinctiveness, not just technical credibility.

Four decisions.

• Repositioned From a Docs Engine to a Trust Layer
  The original product automated docs from pull requests. Once AI made drafting trivial, I repositioned Mari around the emerging risk instead: machine-written content that nobody can verify, source, or approve. It became a single pane of glass over the entire documentation lifecycle, generate, verify, trace, ship, and watch, rather than one more generator.
• A Brutalist Blueprint, Not a Safe SaaS Palette
  The SaaS documentation category defaults to sandy neutrals engineered to feel approachable and read as forgettable. Mari takes the opposite position: a surgical white-and-black blueprint with Bay of Biscay blue for data and Espelette red for alerts, Inter paired with JetBrains Mono, and a node-graph “M” mark. The result is architectural and exacting, and it presents the AI as infrastructure rather than a pseudo-human.
• One Pane of Glass
  The core surface is a documentation-health dashboard. Every AI-written claim is scored and charted with its full history: what is sourced, what is approved, and what has drifted out of date, so a founder reads the health of the whole lifecycle at a glance instead of tab-hopping across tools.
• Made the Invisible Legible
  Each emerging problem gets a live, interactive proof rather than a feature bullet. A verifier intercepts unsupported claims at the gate, a lineage graph traces any claim back to its source and forward to everything that reused it, and a drift monitor flags a doc the instant it falls out of step. The motion is architectural, not decorative: lines that plot themselves in, system data that decrypts into place.

Thesis to shipped.

Not a deck handed to somebody else to build. I owned the product definition and the brand, then wrote the editorial layer that proves the thesis and the documentation that teaches it.

• Product Definition
  Set the thesis and drew the line around it: what Mari is, and which of the many adjacent things it could do were deliberately out of scope. Repositioning from a docs generator to a trust layer meant saying no to most of the roadmap that came before it.
• Brand & Design System
  The Brutalist Blueprint: palette, type pairing, the node-graph mark, motion rules, and the component language, documented as a system a second designer could build in rather than a set of screens to copy.
• The Editorial Layer, Built and Published
  Mari's free plugin is a design system for text: one setup flow, 23 editorial commands, and 171 deterministic detector rules for the tells that AI prose reliably leaves behind. It's open source and installs into any AI coding or writing tool.
• Documentation & Front Door
  Wrote the docs both mari.guru sites are built from, and the landing that has to make an infrastructure argument to a technical reader in about eight seconds.

Taste, made deterministic.

Ask a model to improve prose and you get a different opinion every run. The detector doesn't ask. It's regex, wordlists, density and structural rules, plus small encoder models that run on CPU with no API key and no GPU, and a switch to turn the models off for offline work. Same text, same finding, every time. That's what makes it reviewable. And the vocabulary it hunts isn't a hunch: the post-2023 spike in delve, tapestry and underscore is measured in the literature. Writing editorial judgment down so it survives a model swap is the part that took the taste.