I Taught Claude to Dream About My Book

Last week, Anthropic shipped a feature called Auto Dream.

The pitch is exactly as weird as it sounds. Their AI agents now go “to sleep” between sessions. While they sleep, they review their own memory — every conversation, every decision, every correction the user made — and consolidate it. Resolve contradictions. Prune what’s stale. Re-index what matters.

Modeled, almost on the nose, on what your brain does at night.

The numbers Anthropic published are striking. Big jump in task success rates. Up to six times improvement on long-running projects. All from housekeeping. Same model, cleaner memory, dramatically better output.

I read the announcement and immediately thought: this is the bug that has been quietly killing every long manuscript I have worked on with an AI.

You know the bug. You have hit it.

You are six weeks into a novel. You have made dozens of decisions. Your protagonist’s eyes are hazel, not gray. You cut the sister subplot in chapter four. You moved the inciting incident from spring to autumn. You decided your magic system has a hard cost.

Then you ask Claude for help with chapter twelve. And Claude — confidently, cheerfully, completely — gives you advice based on the previous version of your book. The one you have moved past.

This is not the AI being stupid. It is memory bloat. Your context window has filled up with old, half-baked, contradictory notes from a dozen sessions. The model is doing its best with junk data.

Anthropic’s fix is to let the agent dream. Review. Clean. Re-index.

The catch: their version runs at the agent’s memory layer, on their managed platform, in research preview. Most of you reading this will never touch it.

So I built one we can use. Today. On any Claude session, any manuscript, any book.

I am calling it the Dreaming Skill for Authors.

It is a single skill file plus a small launcher. You drop it into your manuscript folder. You can run it manually inside Claude.ai with zero installation, or set it to run automatically every night while you sleep.

Either way, the skill reads your manuscript, your notes, and a memory file it has been building over time. It pulls out everything durable — your character decisions, your style preferences, your corrections, the rules of your world. It merges what is new with what it already knew. It writes a clean canonical memory file that the AI loads at the start of every future session.

You wake up. Your AI knows your book.

That is the gift. That is what every long-form writer needs and almost nobody has.

The skill runs four phases. Inventory — what files are here? Signal extraction — what is worth remembering? Consolidation — merging the new stuff with what it already knew, resolving contradictions, and pruning what is stale. And writing the outputs — a memory file the AI will load in every future session, plus a dream log entry that tells you what changed.

The dream log is the thing I love most. It is dated, append-only, written in plain English. Last night, I learned X. I reinforced Y. I noticed you have been ending chapters on single-line beats — promoting that to a working pattern. Open question for you: Is Tomas seventy-eight or eighty-two? I have evidence for both.

It reads like a journal of what your AI is learning about your book.

This is not a port of Anthropic’s Auto Dream. Theirs runs at Claude’s memory layer. Mine runs at your project’s memory layer. Theirs is automatic for the AI. Mine is automatic for your book. They are not competing — they are complementary. If Anthropic ever opens up Auto Dream for general use, and they probably will, you should use both.

But until then, this is what we have.

Free. Today. No email gate. No upsell.

It is in my free Ko-fi bundle alongside everything else I have given away this year. The bundle has the skill file, double-click launchers for Mac and Windows, setup guides for nightly automation, integration guides for AuthorClaw and other agents, a visual dashboard, and an example project so you can test it before trusting it on your own book.

Get the free bundle on Ko-fi →

I also recorded a full walkthrough video showing it running on a real manuscript — what the dream log looks like the first morning, how the launcher works, and how the nightly automation gets set up. About six minutes.

Watch the full demo on YouTube →

Why free? Two reasons. Anthropic’s version will (likely) land for everyone eventually. While it is still useful, I want it in your hands. And if you have been on this newsletter a while, you know I do not gate small useful things. Pay for the deeper workflows on Ko-fi if they earn it. The little tools are gifts.

One tactic you can use right now, even without the skill:

Open your current Claude project. At the end of a long working session, type:

Summarize every decision, preference, and correction we made in this conversation that you should remember for future sessions. Format as a brief bulleted list, suitable for a project memory file.

Claude will produce a clean summary. Copy it. Save it somewhere with your manuscript. Next session, paste it in at the top.

That is a manual one-paragraph version of what the full skill does. It will not catch as much. But it will catch the worst stuff. Try it today.

Cheers,

Christopher
Writing Secrets