Second-Brain Auto-Router.

How this fits in the OS.

Why a router, not just a wider capture skill.

Second-Brain Capture is the right tool when you are at your computer and you say "remember this." It is the wrong tool for the inputs you generate when you are not at your computer at all. A voice memo recorded mid-drive, a meeting transcript that finished while you were in the next meeting, a screenshot you saved on your phone six hours ago. None of those moments include a trigger phrase. They only become memory if something else goes looking for them.

The router is that something else. It runs on a schedule. It looks at a small, finite watch list. It classifies new items using the exact same seven buckets the capture skill already uses, so the vault layout stays consistent across both flows. The capture skill is the conscious mouth. The router is the autonomic intake.

What changes for you.

The router is the easy part. The bigger thing is how it changes the relationship between your inputs and your vault. Three shifts you will feel in the first week:

• You stop being the bottleneck. The capture habit only worked when you remembered to invoke it. With the router in front, the inputs you generate while away from the keyboard land in the vault on their own schedule. The friction of capture goes to zero because there is no capture moment to remember.
• You start trusting that your vault is complete. Once a week, you open the dated review folder. If it is empty, the router is doing its job. If there are a handful of misroutes, you tighten one classification rule and move on. The vault stops being a half-honest record of what you decided to save and starts being a near-complete record of what you actually produced.
• You can review a week in one sitting. Because every input lands under the same layout the capture skill uses, weekly review is a single pass across inbox/scratchpad/ and the L1 waypoint files. You see the week without opening the meeting tool, the voice-memo app, or the screenshots folder. The vault is the single surface.

What the router watches, and where it writes.

The router maps source folders to classification buckets. The buckets come straight from the capture skill spec. The router does not invent new ones.

The seven classification buckets the router uses are the seven from the capture skill spec, in the same order: decision, fact or preference, person or relationship, project or build context, article or URL, open question, raw scratchpad. The router does not invent an eighth. If an input genuinely does not fit, it goes to review.

What lands on disk each cycle.

The router writes the same three-layer stack the capture skill writes, with two additions for the non-text inputs. Raw is always written first. Classification can fail; the raw layer cannot.

Before you start.

Four prerequisites. Skip any of them and the router will misbehave on the first real cycle.

• Second-Brain Capture is in place. The capture skill lives at .claude/skills/second-brain-capture/SKILL.md in your workshop repo, your vault skeleton exists at ~/Documents/second-brain/, and you have used the skill at least a few times by hand. The router calls into the skill; if the skill is not working, the router will not work.
• An Obsidian vault exists. Same default location as the capture skill. The L1, L2 Chain of Density Card Catalog Layer, and L3 folder skeleton must exist before the router runs, because the router never invents folders silently.
• A scheduling pattern works on your computer. You have completed Heartbeats & Crons and you have one cron or heartbeat firing reliably. The router is just another scheduled job.
• A transcription tool is installed. MacWhisper running Parakeet is the default. Any Whisper-class transcriber works. The router needs a deterministic command-line call it can run on each new audio file. OCR likewise needs a local or scripted call (Tesseract is fine; macOS Vision via shortcuts works too).

The build.

Seven steps. Capture verification first, watch-list and credentials next, then the router script, then the dry pass, then the cron, then a real cycle. Each step has a duration and a recovery line.

The agent does this for you. Claude reads the foundation skill, drafts the router script, runs the dry pass, and wires the cron. Your job is to pick the watch list, confirm classifications match your intent, and type real values into your password manager when credentials come up.

When it works.

You will see five signals settle into a rhythm during the first week:

1. The cron fires on schedule. Every 15 minutes (or whatever you picked), the router wakes, polls the watch list, and logs how many new items it found. Most cycles will find zero. That is correct.
2. Raw captures land for every input. Open inbox/scratchpad/ at the end of day one. Count the files. Compare against your rough memory of how many voice memos, screenshots, and saved articles you generated. The numbers should match within one or two.
3. The review folder stays small. A handful of items at most after the first week. If the review folder is filling up faster than the scratchpad, your classifier is too cautious; tighten it with Claude.
4. L1 stays tight. Open cairns/L1/work-builds.md and cairns/L1/reading-list.md. The router is gating L1 the same way the capture skill gates L1. Short lists, not walls of one-liners.
5. The diagnostic question. Open a fresh Claude Code session and ask "what did the router catch this week that I would have lost?" If the agent surfaces three or four inputs you genuinely forgot about, the router is earning its place. If the answer is empty, the watch list is too narrow; add a folder.

If you get stuck.

Common failure modes, in rough order of how often they bite:

• The transcription is wrong but the raw audio is preserved. Expected. Whisper-class transcribers make mistakes on accents, technical terms, and quiet audio. The .m4a stays in the source folder; you can re-transcribe later with a stronger model. Do not delete source audio after the router runs.
• The router writes the same file twice. The last-run timestamp got reset or never wrote. Open the watch.json config and confirm the last_run field is updating after each cycle. If it is not, the script is exiting before the timestamp write; have Claude move the timestamp update to the top of the cleanup block.
• The review folder is filling up. The classifier is being too cautious. Quote the capture skill spec back at Claude and ask it to tighten the prompt. The seven buckets cover almost every real input; very few things should land in review.
• The cron fires but nothing happens. The router cannot find the watch folders. Most common on a new computer or after moving Documents. Run the script by hand with verbose logging and check the path errors before debugging anything else.
• A real credential ended up in chat. Treat the key as compromised. Rotate it in the source service first. Then rebuild the entry in Infisical or 1Password with PASTE_HERE and type the new value in yourself. If Claude offers to enter the real value, say no.
• An L2 Chain of Density Card Catalog Layer card got written for something that should have stayed scratchpad. The router is over-promoting. The cairns-ingest handoff should only fire for source-worthy items (full articles, substantial transcripts, real meetings). Open the router script and have Claude tighten the source-worthy threshold.
• The router stops firing after a few days. Heartbeat or cron drift. Same recovery as the Morning Brief: confirm the cron entry survived, confirm the computer was awake during the missed firings, confirm no permissions popped up that paused the runtime.

Still stuck? Post the cycle's log output, the watch.json config (with paths but no secrets), and the input that misrouted. Someone in the cohort has hit the same shape.

Where to go next.

The router is the autonomic intake. The next moves on the shelf either feed it more sources, give it richer downstream processing, or wire it into the rest of the production stack.