← All cards

Writing & talks

From connected cards to an atomic content studio

A new authoring model for this site

The first version of this website’s content studio made the Jekyll files visible through a local interface. That was useful, but it also exposed a deeper problem: the folders and editing tools did not yet express the content model clearly enough.

The new direction begins with a simple distinction. Cards are atomic notes. Pages are structures.

Cards have one centre of gravity

An atomic card is not necessarily short. It holds one idea, experience, question, or reusable explanation that can stand alone and connect to other cards. Context and examples belong when they help that idea; a second independently useful idea may deserve its own card.

This gives the studio a practical editorial question: is this one card, several cards that should be split, or duplicate cards that should be merged? The studio can now flag possible split and merge opportunities as proposals. It does not silently rewrite them. Related but distinct ideas should stay separate and gain an explained connection.

Pages are allowed to compose

The homepage, category directory, card directory, and changelog do a different job. They arrange many things, sometimes without claiming that every item is semantically related. Pages can therefore use structural HTML, Liquid includes, and loops when a layout needs them.

Cards are Markdown-first. The local editor offers a simple writing view, Markdown source, and the rendered Jekyll preview. Existing cards with hand-built HTML remain editable in source mode until they are deliberately converted; the editor does not risk a lossy automatic rewrite.

Internal links can still use Jekyll’s template tags. A Markdown link such as [Connected notes]({% include card-url.html slug='connected-notes' %}) resolves through the card’s stable identity. The studio’s searchable card picker can insert that syntax without asking an author to remember it.

Search is part of authoring

The same search-and-filter behaviour now supports navigating content, finding a group, choosing a relationship target, and spotting a title that may already exist. Creating a group is explicit: if the search finds no active group, the author can choose to add it rather than accidentally creating one through a typo.

Suggestions should explain themselves

Link suggestions begin with an inspectable deterministic ranking. Title words receive more weight than descriptive prose, which receives more weight than the body. The proposal shows lexical similarity, shared-term coverage, common graph neighbours, group affinity, and the terms that actually matched.

That number is a rank, not a declaration of truth. It helps decide what to inspect next. The author still chooses whether the connection exists, which relationship type describes it, and what evidence-bearing reason belongs with it.

Codex can review the deterministic shortlist with either its cloud model or a model served locally through Ollama. Both routes use the same agent runtime, repository instructions, read-only evidence tools, and structured output contract. Every model call is manual: no model runs while opening, typing, saving, building, or watching files. Its output becomes a proposal to accept or dismiss.

Proposals can now answer questions

A proposal is easier to judge when I can ask what it means, where its evidence came from, or whether the site already contains a related card. Each proposal therefore has a private, persistent conversation. The agent can search and retrieve cards, pages, relationships, and graph context, but it cannot run commands, edit files, change the proposal, or apply it.

These conversations use the proposal’s original cloud or local model when that information exists. Deterministic proposals begin with the configured local model. Each turn is captured in the private audit inspector, including the provider, resolved model, evidence activity, validation result, and sanitized agent events.

The response boundary is deliberately strict. Relevant source material is supplied before inference, and an answer that merely says “let me check” or describes a future search is rejected. The agent gets one corrective attempt to finish the retrieval and answer the question directly.

Proposals now sit with the work they concern: link proposals under link suggestions, split and merge proposals under atomicity review, and content changes beside the editor. The side panels can be resized, while the link picker has moved into the editor tools and keeps Markdown visibly in sync when it inserts a link.

The files remain the source of truth

The studio stores local preferences—feature flags, panel widths, endpoint, and model choices—but not API keys. Codex uses its existing login, local inference stays on a loopback ODS or Ollama endpoint, and a diagnostics button checks each setup only when asked. Available local models come from the configured endpoint rather than from a free-text setting.

Every explicit inference attempt also has a durable private audit trace. Successful answers, provider failures, timeouts, malformed output, validation failures, model discovery, and tool events can be inspected without turning operational data into published or canonical content.

The important boundary has not changed: Markdown, YAML, Liquid, Jekyll, and Git remain the durable system. The studio is becoming a more legible way to work with that system, not a second database sitting in front of it.

Immediate connections

Explore the cards directly connected to this one.

Connected cards

Follow links in either direction, or find cards that mention this one.

Show