Cairns Ingest Pipeline.

How this fits in the OS.

What you'll build.

The ingest pipeline is the entry point to your second brain. If Librarian is how knowledge comes out of Cairns, ingest is how knowledge goes in. One skill takes any source you point it at, classifies the type, writes the raw evidence to L3 first, runs a chain-of-density pass to produce an L2 Chain of Density Card Catalog Layer card, and only then proposes whether the source should change your always-on L1 context.

The whole point of this build is that the agent does the work. You describe what you want to remember in plain English. The agent classifies, writes, summarizes, links, and proposes. You approve each step. The structural discipline of the pipeline (raw preserved first, dense cards cite raw, L1 never silently updated) is enforced in the skill itself, not left to your willpower.

What changes for you.

The nine-step flow is the easy part. The bigger thing is the shift in how you treat capture. Three internal shifts you'll feel:

• Sources stop evaporating. You read an article worth remembering, you hand it to /cairns-ingest, and it lands in your vault with a dense card and a backlink to the raw. You stop bookmarking and forgetting.
• Raw evidence becomes load-bearing, not optional. The pipeline writes L3 before it writes anything else. Your dense summary is never the only record. Six months later when you want to audit what was actually said, the verbatim is there.
• Your standing context stays curated. The agent doesn't write to L1 silently. It proposes a waypoint only when a source changes how you operate, and asks before touching the always-on layer. Memory drift gets gated at the door.

Before you start.

Two prerequisites:

• Cairns is set up. You've worked through The Cairns Three Tiers and you have at least the L1, L2 Chain of Density Card Catalog Layer, and L3 folder skeleton in your vault. Ingest writes into folders that have to exist.
• The skill is installed. The cairns-ingest skill ships in the workshop repo at .claude/skills/cairns-ingest/SKILL.md. Once it's there, Claude Code picks it up automatically when you ask to ingest, capture, add, or process a source.

No credentials, no API keys, no external services. This is local markdown only.

Walk through the build.

Why this shape, not a bigger one.

The ingest pipeline looks minimal. Raw plus dense card plus optional waypoint. That's by design. The shape works at small corpus sizes without fancy retrieval, and it scales by adding layers later instead of rewriting the foundation.

Karpathy also flagged the write-side complement that the ingest pipeline alone doesn't solve: outputs of agent queries should be filed back into the wiki, so each exploration enhances retrieval for the next one. That loop closes elsewhere in the OS, on the consolidation passes and the decision logger. Ingest handles the input side. Closing the loop is on the shelf alongside it.

When it works.

You'll see three things land in your vault on a successful ingest:

• A new file under cairns/L3/<type>/ with the complete source, verbatim. Open it. The content should match the source exactly.
• A new file under cairns/L2/cards/ with the dense 120 to 180 word summary, the retrieval hooks, the key entities, and the backlink to the L3 file. Open it. The summary should feel tight, not fluffy.
• Either an L1 proposal you approved (one new line in the right waypoint file) or no L1 change at all. The default outcome is no L1 change. That's correct. Most sources don't shift standing context.

The compounding test: ingest five sources over a week. At the end, ask your agent "summarize what I've captured this week and surface anything that connects." If the agent answers by scanning the L2 Chain of Density Card Catalog Layer cards (not grepping raw), the pipeline is doing its job.

If you get stuck.

Common failure modes, ranked by how often they bite people:

• Vault path was ambiguous and the skill guessed wrong. Most common on the first run. The skill defaults to ${HOME}/Documents/second-brain on macOS and the user's Documents folder on Windows. If the default is wrong for you, tell the agent explicitly: "the vault root is <path>, use that for every cairns-ingest from now on."
• The L1, L2 Chain of Density Card Catalog Layer, or L3 folders didn't exist yet. The atomic write needs targets. If the folders are missing, the skill can create them, but it's cleaner to set up your Cairns skeleton first via the Cairns Three Tiers card.
• The L2 Chain of Density Card Catalog Layer card lost something load-bearing. Chain of density is lossy by design. The read-back step is your moment to catch a missing entity or a softened claim. Edit the L2 Chain of Density Card Catalog Layer card in place. Never edit the L3 raw; that's the evidence layer.
• The agent tried to write to L1 silently. If you see an L1 file change you didn't approve, stop and ask the agent which step skipped the proposal gate. The spec is explicit: never silently update L1. If a session ever does, treat that as a bug in the skill setup, not a feature.
• You ingested a source that was actually a decision. The skill should redirect you to /log-decision when it detects decision-shaped input. If it doesn't, redirect yourself. Decisions belong under cairns/L2/decisions/ with the atomic five-file write, not under cairns/L2/cards/.
• A write failed mid-run and you're not sure about rollback. The spec says rollback is automatic; if rollback itself fails, the skill stops and tells you the exact paths needing manual review. If you ever see a partial ingest, do not re-run the skill on the same source until you've cleaned up the partial files using the paths the agent reported.

Still stuck? Post the prompt, the agent's response, and what you expected, in the workshop Slack. Someone in the cohort is solving the same thing.

Where to go next.

You ran the pipeline end to end on one source. The compounding starts when the pattern becomes habit. The natural next moves on the shelf: