Chain-of-Density Cards.
The compression technique behind every Cairns card. Five rewrite passes, each one denser than the last, evidence preserved. The card stays short. The signal goes up.
- The five-pass rewrite loop that turns a vague first-draft summary into a dense, entity-rich card the agent can actually use to decide what to fetch.
- A worked example showing the difference between a weak L2 Chain of Density Card Catalog Layer (Card Catalog Layer for short) card and a strong one, side by side.
- The skill that runs the loop for you Hand a source to /cairns-ingest. The agent does the rewriting and proposes the final card. You approve, edit, or skip.
Run the loop on a real source.
Pick something you actually want to remember: a transcript, an article, a meeting note. Hand it to the agent and watch the five passes happen.
How this fits in the OS.
The Cairns Three Tiers. Chain-of-Density is the technique that builds the middle layer (L2 Chain of Density Card Catalog Layer (Card Catalog Layer for short) cards) from raw sources sitting in L3.
A source worth preserving (transcript, article, meeting, paper, Slack thread), the /cairns-ingest skill, and a human approval step before the card lands in your vault.
A Card Catalog Layer the agent can actually navigate. Source-backed answers. The ingest pipeline that turns new material into structured memory automatically.
The shape of the problem.
A vague summary is worse than no summary. If your L2 Chain of Density Card Catalog Layer card says "this video talks about AI tools and productivity," the agent learns nothing from reading it. It still has to load the full transcript to decide whether the source is relevant. The card adds a step without adding signal.
The fix is not to write longer summaries. Longer summaries crowd context and still miss the specifics that make a source useful. The fix is to write the same length summary more densely: pack in the real entities, claims, decisions, and constraints from the source, and cut everything that doesn't earn its place.
That is the job Chain-of-Density does. It is a compression technique, not a summarization technique. The difference matters.
A summary tries to be readable on its own. A dense card tries to be useful to an agent deciding whether to fetch the source. Different goal, different writing. The card is allowed to read a little tight. It is not a blog post. It is a library card.
The five-pass loop.
The technical name is Chain-of-Density. In plain English: write a draft, find what you missed, rewrite the same length with more signal, repeat. Five passes total. The length stays roughly constant. The information density goes up each round.
Sparse summary
Write a basic 80 to 120 word summary of the source. Don't try to be dense yet. Just get the shape down.
Missing entities
List 5 to 10 important things the summary doesn't yet mention: names, tools, concepts, claims, dates, constraints. This is the audit step. Most cards fail here because the writer never looks for what got cut.
Denser rewrite
Rewrite at the same length, but fold the missing entities back in. Something else has to give to make room. Cut filler phrases, hedging, restatements, scene-setting that doesn't carry signal.
Missing retrieval hooks
List the questions this source should answer that the current draft doesn't support. A retrieval hook is the kind of phrase a future-you might type into the agent. If the card doesn't contain language that would surface it, the agent won't find it.
Final dense card
One last rewrite at the same length. Bake the retrieval hooks into the prose. No generic phrases like "the video discusses" unless they carry meaning. The goal is not a longer summary. The goal is a better card.
You don't sit there counting words and listing missing entities by hand. Hand a source to /cairns-ingest and the skill runs the five passes, shows you each one, and proposes the final card. You read it, approve, edit, or rerun. Nothing lands in your vault without your sign-off.
What this looks like in practice.
The same source, summarized two different ways. The first is the kind of card that gets written when no one runs the loop. The second is what comes out of pass 4.
"This video talks about AI tools and productivity. It covers workflows, automation, and different ways to use agents."
"Cole Medin demonstrates a Claude Code second-brain workflow that stores YouTube transcripts as raw evidence, creates Obsidian-readable summaries, and routes retrieval through a local agent before escalating to vector search. The useful pattern is not summarize videos; it is source-backed ingestion: raw transcript preserved, dense card generated, then agent uses the card to decide whether to load the transcript."
Same length range. The first names no source-specific signal. The second names the creator, the tool, the storage pattern, the routing logic, and the underlying principle. An agent reading the second one can decide whether to fetch the L3 transcript. An agent reading the first one cannot.
Why evidence preservation matters.
Every L2 Chain of Density Card Catalog Layer card points back to the raw source it was distilled from. That backlink is not optional. It is what makes the card trustworthy.
The agent can answer most questions from the card alone. But the moment a question needs verification, a quote, a number, a specific claim, the card is just a pointer. The agent follows the backlink down to L3 and reads the original. The card was never trying to replace the source. It was trying to be the scan target that decides whether the source is worth opening.
This is why the card stays short. A 400-word card is no longer a card. It is starting to compete with the source it points to. Discipline the length. Trust the link.
The research the technique is named after (arxiv.org/abs/2309.04269) showed that summaries get better when you repeatedly add missing salient entities while holding length fixed. The model's first pass was readable but vague. Each subsequent pass packed in more entities at the same word count. The final summaries were rated more informative without becoming harder to read. The Cairns adaptation is a practical version of that loop, with retrieval hooks added in pass 3 because cards have a job summaries don't: they have to be findable, not just informative.
The discipline that makes it work.
Three rules keep your Card Catalog Layer useful as it grows:
- One source, one card. Don't create a "summary of all AI videos" card. That is too broad to retrieve well. One transcript, one paper, one meeting, one Slack thread worth keeping. Each gets its own card.
- Cards are append-only after first write. If new context changes how you read a source, write a new dated card with a backlink to the prior version. Don't edit in place. The audit trail matters.
- Every card has a promotion decision. At the end of the card, one sentence: does this source change your standing context enough to update an L1 waypoint, or not? Most of the time the answer is no. Writing it down forces the decision instead of letting cards accidentally promote themselves.
You don't enforce these by hand either. /cairns-lint audits your vault for weak cards, broken backlinks, and stale waypoints. Run it when something feels off. The discipline is on the rails, not in your head.
Where to go next.
You understand the compression technique. The natural next moves: