Cairns Dream.
Sleep mode for your second brain. End-of-day pass distills the day's activity into cards and proposes memory updates. Never auto-writes.
- An overnight pass that reads the day's raw activity (files touched, commits, conversations, notes) and proposes dense summary cards for your memory tier.
- A morning briefing waiting when you sit down: N cards drafted, N waypoints proposed, top three themes, and anything surprising that crossed a frequency threshold.
- Human gates on every write the agent drafts. You approve, edit, or skip. Nothing crosses into permanent memory without your sign-off, by design.
Wire up your first dream pass.
Hand the spec to Claude in one paste. You will install the slash command, run a session-scoped pass live, then schedule a nightly run that writes its proposals to a file you read in the morning.
How this fits in the OS.
The Cairns Three Tiers already in place. You need L1 waypoints (the always-on pins), an L2 Chain of Density Card Catalog Layer (Card Catalog Layer for short) (dense summaries), and L3 raw activity (files, commits, notes) for the dream pass to have anything to consolidate. A working scheduler on your computer for the nightly run.
One slash command (/cairns-dream), the Chain of Density technique for distillation, and cron or Task Scheduler for the overnight trigger. Optionally a secrets manager so the scheduled job can sign in without a human present.
The shift from "I need to write this down" to "the agent already wrote it down, I just have to read what it drafted." Capture stops being a task. Approval becomes the only thing you do by hand.
What you will build.
A slash command called /cairns-dream that runs a memory consolidation pass. The metaphor is biological sleep. During the day you accumulate raw activity. Overnight, the agent reads what changed, drafts dense summary cards (Chain of Density technique, three to five passes, 120 to 180 words each), and proposes updates to your always-on waypoints.
The pass has eight steps, and a human approval gate sits on every write. The agent never auto-writes to memory. You wake up, read the report, and say approve, edit, or skip on each proposal. What survives gets pinned. What doesn't, doesn't.
Five scopes are supported. Pick the smallest one that fits the cadence you want to run.
What changes for you.
You stop being the consolidation layer. Three shifts you will feel inside the first week:
- You wake up to a consolidated yesterday. The morning briefing is already written. You open the report, gate the proposals, and you are done with capture for the day. The work that used to live in your head as "I need to write that down later" is drafted and waiting.
- Patterns surface that you would have missed. The pass flags anything you referenced three or more times during the period. If you mentioned a topic five times this week and it isn't in any L1 category, the agent surfaces that as a candidate. You decide whether it earns a pin.
- Approval becomes the only manual move. Drafting cards, scanning for themes, checking for repeats, suggesting waypoint edits, all of it runs while you sleep. The only thing you do by hand is read three lines per proposal and say approve, edit, or skip.
Before you start.
This is an 8D build sitting on top of two earlier shelves. Make sure these are in place before you wire up the dream pass:
- The Cairns Three Tiers. L1 waypoints, L2 Chain of Density Card Catalog Layer, L3 raw, all configured against your knowledge base. Dream is a consolidation pass over that structure. Without it, there is nothing to consolidate into.
- A working scheduler. Cron on Mac or Linux. Task Scheduler on Windows, pointed at Git Bash. You will use it at the end to fire the nightly pass.
/personalizedone. The pass leans on the personalization flow's understanding of your projects, your categories, and what you actually work on. Run that first if you haven't.- An L3 source the agent can read. Git history is the easiest one (commit messages, file diffs). Modified files in your knowledge folder also work. The richer the L3 raw, the better the L2 Chain of Density Card Catalog Layer drafts will be.
The build.
Six steps, each one running inside the same starter prompt above. The agent does the work. You confirm between steps.
Install the slash command
The canonical spec ships in the workshop repo at configs/commands/cairns-dream.md. Drop it into your project's slash command directory (or your global ~/.claude/commands/ if you want it available everywhere). The file has the full eight-step procedure baked in. Claude reads it when you invoke /cairns-dream.
The agent does this for you. Hand it the path, it copies the file, confirms it loads.
Run a live session-scoped pass
Before scheduling anything, run the pass live against the current conversation so you can see all eight steps. Invoke /cairns-dream --scope=this-session. The agent will:
- Confirm the scope.
- Read the relevant L3 raw (this conversation).
- Distill into L2 Chain of Density Card Catalog Layer card candidates. One source per card. 120 to 180 word dense summaries after three to five density passes. Specific facts, decisions, retrieval hooks, backlinks to L3.
- Propose L1 waypoint updates if anything was referenced three or more times.
- Show you each L2 Chain of Density Card Catalog Layer card draft and ask approveeditskip
- Apply approved L1 updates.
- Print the dream report.
- Offer to schedule the nightly run.
The agent does this for you. Drafting, distilling, theme detection, surprise detection. Your job is the gate.
Approve, edit, or skip every proposal
This is the human-in-the-loop core of the build. The slash command spec is explicit: nothing auto-writes. Every L2 Chain of Density Card Catalog Layer card and every L1 waypoint update is gated.
- Approve writes the card to
L2/cards/YYYY-MM-DD-<slug>.mdand applies the waypoint update. - Edit takes your revision and then writes.
- Skip discards the candidate.
The gate exists for a reason. The agent is doing a Chain of Density pass on raw activity, and Chain of Density is opinionated. The pass will sometimes propose a card that doesn't match how you actually think about the source. Skip those. The ones that do match are dense, well-cited, and ready to live in your memory tier without further work.
Read the dream report
The pass ends with a summary that tells you what just happened. You'll see something like this:
The "surprising" line is the most valuable output. It catches frequency patterns you would not have noticed yourself. If a topic crosses the threshold and isn't pinned, that is signal.
Schedule the nightly run
Once you have run a live pass and you trust the proposal quality, schedule the daily version. The recommended default is /cairns-dream --scope=today at 03:00, with the report written to a file you will read in the morning instead of streamed to a chat window.
On Mac or Linux, that is a single cron entry. On Windows, a Task Scheduler job pointed at Git Bash running the same shell command. The Heartbeats and Crons card walks the install end to end.
The agent does this for you. It writes the shell script, registers the schedule, verifies the next run, and leaves you with the one-liner you can use to check the log later.
Build the morning review habit
The dream pass is only as useful as your habit of reading what it drafted. Most builds fail here, not in the install.
Add the report file path to whatever you open first in the morning. If you use a morning briefing slash command, have it include the dream report in its output. The proposals expire as new ones land, so a daily two-minute review keeps the queue from piling up.
The pattern is consumer-grade overnight consolidation. The team's production version (the metalog of evals, the queue that fans out diffs to an embedding pass, the deferred-fix loop where recurring flukes get promoted to bugs and overnight agents update the rule sources) is the same shape running on production infrastructure. /cairns-dream is the version you run on your own computer with one human gate.
When it works.
Three signals you have the pass dialed in:
- The morning briefing has fewer than ten proposals. If the count keeps climbing, the scope is too wide, or the pass isn't running often enough. A nightly
--scope=todayon a daily writer should land around three to seven cards plus one or two waypoint updates. Larger than that means yesterday was unusually heavy, not that the pass is broken. - You approve most cards without edits. If you find yourself rewriting every proposal, the L3 raw is thin or the density passes aren't running. Check the spec, the Chain of Density step is the load-bearing one. If you find yourself skipping most proposals, the scope is dragging in noise. Narrow it.
- The "surprising" line earns its keep. Roughly once a week, the frequency-detection pass surfaces something you would not have flagged yourself, and you decide to promote it. That is the pattern working. If the surprising line is empty for two weeks in a row, the pass is being too conservative or your activity is too narrow to surface cross-source patterns.
If you get stuck.
Common failure modes, ranked by how often they bite:
- No L3 raw to read. The pass scans git logs, modified files, recent transcripts. If your knowledge work doesn't leave a file trail, dream has nothing to consolidate. Fix by routing capture (transcripts, daily notes, decision logs) into a folder the pass can read.
- The pass auto-wrote something. It shouldn't. The spec is explicit that nothing crosses into permanent memory without a human gate. If you see writes appearing without approval prompts, the slash command file got mangled in install. Re-copy from
configs/commands/cairns-dream.mdand run again. - The cards are too vague. Either the L3 raw was thin, or the density passes didn't run. A good L2 Chain of Density Card Catalog Layer card has specific facts, named entities, decisions with dates, and backlinks. If your draft is "we discussed scheduling," that is a pre-density draft. Push back, ask for another pass.
- The proposal queue piles up. You scheduled the pass but you aren't reading the report. Either move the morning review to a place you cannot skip (start of your morning briefing, top of your daily note template) or reduce the cadence to every other day until you have the habit.
- Cron fires but nothing happens. Scheduled jobs run in a different shell environment than your interactive one. The script needs to load your secrets and your PATH explicitly. The Heartbeats and Crons card has the install pattern that handles this. Don't debug it from scratch.
Where to go next.
Dream is the consolidation pass on top of the Cairns stack. The natural next moves once it is running: