Second-Brain Capture.

How this fits in the OS.

Why a capture skill, not a slash command.

The capture problem is not "I don't have a place to put things." Most people have ten places. A Notes app, a Slack DM to themselves, a voicemail to themselves, a Google Doc, a sticky note on the monitor, a paragraph at the bottom of a meeting transcript. The problem is the friction of choosing which one and the certainty that you will never look at most of them again.

A slash command makes you remember to type the slash. A skill triggers on natural language. You say "remember this" or "save this", drop the input, and the agent does the routing. The skill exists for the moments you are mid-thought and do not want to break flow to think about filing.

When to use which intake path

The capture skill is the wide mouth. It absorbs anything. But a few inputs have sharper tools that should win when they apply:

• You just made a decision. Use /log-decision directly. It writes the same five-file atomic structure the capture flow would hand off to anyway, with no intermediate scratchpad step.
• You are ingesting a substantial source (a full article, a transcript, a paper, a meeting recording). Use cairns-ingest directly. The capture skill will hand off to it anyway, so you save a hop.
• You have a loose thought, a link, a screenshot of someone's tweet, a voice memo about something you read. This is exactly what the capture skill is for. Let it classify and route.

What changes for you.

The skill is the easy part. The bigger thing is how it changes the moment a thought arrives. Three shifts you will feel:

• You stop choosing where things go. "Remember this" is the only phrase you need. Classification happens after the fact, by the agent, with one clarifying question if the input is genuinely ambiguous. You return to your work in seconds, not minutes.
• Raw stays raw. The skill writes your verbatim input first, before any summarization. If the dense card later misreads your meaning, the original is one folder away. This is the same epistemic posture immutable transcripts have in the workshop's Cairns architecture, applied to live capture.
• The L1 layer stops drowning. Not every captured thought gets promoted to a waypoint. The skill is explicit about this. Only items that are L1-worthy (a decision in flight, a new person, a build context shift, a reading-list addition) get a one-line waypoint. The rest stay parked in the scratchpad, dated, where the agent can still find them but the always-loaded context does not get flooded.

What the skill routes, and where.

The skill classifies every input into one of seven buckets and writes to a specific path for each. Straight from the skill spec:

The classification is one short prompt, not a multi-step interrogation. If the input is genuinely ambiguous, the skill asks one follow-up question and then moves on. The whole point is that capture should feel like dropping something on a desk, not filling out a form.

What lands on disk.

Every capture writes at least one file. Source-worthy captures can write up to three. The raw file is always written first.

The skill then reports back: raw path, L2 Chain of Density Card Catalog Layer card path if created, L1 file updated if any, follow-up question if needed. No explanation, no preamble. You see where it went and whether anything is blocked.

Before you start.

Three prerequisites, all quick:

• A second-brain vault exists on your computer. Default location is ~/Documents/second-brain. On Windows, that resolves to your Documents folder. If you haven't created it yet, the build prompt below has the agent create it for you on first run.
• 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. Capture writes into folders that have to exist.
• An Obsidian (or equivalent) viewer is open on the vault. Not required for capture to work, but the moment you see your first capture show up in the file tree, the loop closes and the habit starts forming.

Walk through the build.

When it works.

You'll see four things settle into a rhythm:

• Your scratchpad fills up. Open inbox/scratchpad/ a week in. The file count is the visible measure of capture friction going to zero. If it's empty, the trigger phrases aren't entering your vocabulary yet; force a few in a row deliberately.
• Your L1 waypoint files stay tight. Open cairns/L1/reading-list.md or cairns/L1/work-builds.md. You should see a clean short list, not a wall of one-liners. The skill is gating L1 the way the spec says it should.
• Source-worthy items have L2 Chain of Density Card Catalog Layer cards. The full article you captured on day three should have a card in cairns/L2/cards/ with the core claim, retrieval hooks, and a pointer back to the raw scratchpad. The loose thoughts should not.
• The diagnostic question. Open a fresh Claude Code session and ask "what did I capture this week?" If the agent surfaces your captures by reading L1 plus scanning the scratchpad, the architecture is wired. If it has to grep the filesystem, something is not landing where the spec says.

If you get stuck.

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

• You stop using the trigger phrases. Most common. The skill only fires on natural-language triggers like "remember this", "save this", "capture this", or "add this to my second brain." If you slip back into ad-hoc note-taking, the skill never runs. Pin the triggers to a sticky note for the first two weeks.
• The vault path was ambiguous and the agent guessed wrong. Happens on a fresh computer or after moving your Documents folder. The fix: open the skill spec and confirm the default path. If it doesn't match where you want captures to land, tell the agent explicitly: "the vault root is <path>, use that for every capture from now on."
• Every capture is becoming an L2 Chain of Density Card Catalog Layer card. The skill is over-promoting. The spec says enrich only when useful and update L1 only when warranted. Read it back to Claude and ask it to be stricter about gating. Loose thoughts should land in scratchpad and stop there.
• Captures contain decisions that should have used /log-decision. The skill is supposed to offer the decision flow when it spots one. If it isn't, paste the spec back in chat and ask the agent to re-read the classification rules. A decision logged through capture instead of through /log-decision ends up in scratchpad with no atomic five-file write, which is the wrong outcome.
• The L3 raw file got summarized. Never acceptable. The raw layer is the evidence layer; it stays verbatim. If you see summarization where the spec says "Raw", report it and re-run with the spec quoted back at the agent. Edit only the Initial Read section; never the Raw section.

Still stuck? Post the trigger phrase you used, the input you dropped, and what the skill did with it, in the workshop Slack. Someone in the cohort is solving the same thing.

Where to go next.

You wired up the intake mouth. The natural next moves on the shelf: