Lossless Changelog
product

Two Doors, One Room — The How-To Engine Starts, and the Registry Grows

We started a documentation engine that the Dojo actually needs — how-to content you can enter two ways: by the problem you're solving (a use-case like 'automate dealflow') or by the tool you've heard you should learn (a guide like 'Hermes Agent'). Same content, two doors, cross-linked. It's built on @lossless-group/lfm (not Starlight), shipped a first wired slice end-to-end, gained internal /tools pages that wikilinks now prefer over external sites, and a /recipes index in the header. Alongside it, the tool + participant registries grew: Composio and Granola joined, and David Rolf, Abhinaya Konduru, and a fixed-up Lylan Masterman landed with real profiles.

Michael StatonMusic Editoraugmented with Claude Code on Opus 4.8 (1M context)

Two Doors, One Room — The How-To Engine Starts, and the Registry Grows

Why Care?

VCs arrive at AI two ways, often in the same sentence: "I need to automate my dealflow" (a problem) and "I've heard I need to use Hermes Agent" (a tool). A docs site that only indexes one way fails half the room. So we started building one that indexes both — the same how-to content reachable from the use-case door and the tool door, cross-linked so you can step between them.

If you've ever bounced off a tool's docs because you didn't know why you'd use it, or bounced off a "10 AI tools for VCs" listicle because none told you how — this is the thing that sits in the middle and answers both. It's the start of a real knowledge base for the Dojo, built from our own community's named asks.

What's New?

The how-to engine (new):

  • A spec and a working first slice. [[context-v/specs/Two-Perspective-How-To-Docs-Engine-on-LFM.md|The spec]] lays out a "two doors, one room" model; the first vertical slice wires the Automate Dealflow use-case to the Put Hermes Agent on Your Dealflow guide, both ways.

  • /recipes in the header — an umbrella index listing both perspectives as cards: Use-Cases (the problem) and Guides (the how-to), with draft badges and tool chips.

  • Internal /tools/<handle> pages — every registry tool now has its own page (info + body + the guides, use-cases, and people that reference it). Wikilinks prefer our page over the tool's external homepage now (so [[tools/hermes-agent]] lands on FullStack VC, not nousresearch.com).

  • Live LFM rendering — callouts, [[wikilinks]], and the rest render through @lossless-group/lfm and our AstroMarkdown component. No Starlight — we built the routes ourselves so the two-perspective model isn't fighting a single-tree framework.

The registry grew up:

  • Two tools: [[tools/composio|Composio]] (the tool-calling + auth layer for agents) and [[tools/granola|Granola]] (the AI meeting notepad — the benchmark the open-source crowd is cloning).

  • Two new fellows, fully sourced: [[davidrolf|David Rolf]] (Head of Visa Ventures, Class 12) and [[abhinayakonduru|Abhinaya Konduru]] (Partner at M25, Class 28) — real bios, local headshots, and their stated asks captured.

  • A stub repaired: [[lylanm|Lylan Masterman]] went from name: lylanm + an expiring LinkedIn-CDN avatar to a real profile (Co-Founder & Managing Partner, Whipsmart Ventures) with a committed headshot.

  • Toby's June 24 demo reframed from a laundry list into a coherent "everything funnels into one second brain" arc.

The Story

The naming nearly tied us in knots. The first cut had three buckets — use-cases, tool-uses, and "recipes." A look at it in plain language settled it: the two doors that read well as URLs are /use-cases/ and /guides/, the tool angle rides on the tools registry we already have, and "recipe" is just the navigation idea — a guide surfaced from both sides. We wrote that down as provisional and, crucially, resolved it by authoring real content instead of arguing on paper. One use-case, one guide, wired both ways — and the structure fell out exactly as hoped.

The most satisfying small win: the wikilink resolver. The first version sent [[tools/hermes-agent]] to Nous Research's homepage. That's backwards — we want readers on our site, where the tool page shows which guides and people actually use it. Flipping the resolver to prefer an internal page, fall back to external turned a leaky link into a loop: use-case → guide → tool → back to the guide.

The registry work was the quieter half of the day, but it's the same instinct: the survey and the session intros keep naming tools and people, so we keep making them legible. David Rolf showed up with a perfect testimonial ("I haven't seen a straightforward way to put it to use… 'putting it all together' sounds VERY compelling") — which is, almost word for word, the problem the how-to engine exists to solve.

Under the Hood

  • Soft collections. use-cases and guides are glob() collections with .passthrough() schemas — most fields optional, never hard-fail on extra frontmatter (the YAML-leniency house rule). A guide carries its relations (use_cases: / tools:) in frontmatter; both perspective lists are derived from those, never hand-mirrored.

  • The resolver (src/lib/docs/wikilink-resolver.ts) maps use-cases/, guides/, people/, bare handles → /people/<handle>, and tools/<handle> → an internal /tools/<handle> page when the handle is in the registry (external url only as fallback).

  • Honest drafts. The Hermes guide is status: Draft / publish: false with > [!todo] callouts where tool-specific steps need a real run to verify — the framing is grounded in what Hermes is, the exact commands are flagged, not fabricated.

  • No new type debt. Every step held astro check at the pre-existing baseline (16 errors, all in older files); the new routes carry zero.

What's Next

The content layer and a raw, themed read are live. The docs chrome isn't yet: left-hand navigation, a right-hand table of contents, and prev/next pagination are specced and waiting. And the engine wants more content to pressure-test it — the survey's other named asks (accelerate memo drafts, wire Claude into the VC stack) are the obvious next guides.