Brain: ask what exists, paste anything, see where it fits
The Brain is the first of the three sections, before Plans and Building. Plans is one piece of work in detail; Building is agents that can write. The Brain is the whole project with nothing committed: what is already built, the ideas you have not decided on, where each one fits, and whether it already exists. Nothing in it can change your code.
Switch to it with the control at the top of the agent list, or press ⌘⇧P to cycle Brain → Plans → Building.
The Brain was called Product until November 2026. Chats, tabs and links from before then open here; /docs/product comes to this page.
Ask
Ask continues the project's Brain chat: "is CSV export built?", "where does billing fit?", "what is on other branches?". Each project has one, so the conversation carries on where you left it; New brain chat starts a fresh thread that becomes the one Ask continues.
The Brain reads the product map, your ideas and plans, the Changes log, its own memory, the project's branches and its files, and answers in four scopes: in this checkout (naming the files), on another branch (naming the branch), live on a target (only when a change or release record says so) or unknown. When it cannot find something it says what it checked — never that something "does not exist".
When something is not built yet it sizes it and offers six choices: keep it as an idea, plan it now, split it into several plans, start building now, make it a direction for future plans, or skip it and remember why. It records your pick, and the pane beside the chat shows what it recorded.
Refresh the brain reads the project's own docs (AGENTS.md first, then any brain folder, the README and docs), the map, plans, the Changes log, releases, issues and branches, and records what it found as facts with evidence. It is an ordinary read-only chat you can open, with its cost on the Usage page (Included on your Claude plan, Cash with an API key). It decides nothing and cannot change your code. It records what it read and what it could not — a source it could not reach is shown as unknown, never as "not built". A refresh that records nothing counts as failed. Choose when it runs: by hand, after each build (once, ten minutes after the last one), or daily at an hour; a refresh that could not run is listed with the reason.
Changing your docs. When a conversation shows that AGENTS.md, the README or a doc is out of date, the Brain proposes the change as a diff; nothing is written until you apply it, and it applies only if the file has not changed since. It never touches the project's .nocturnal/ folder that way, and AGENTS.md only when you allow Nocturnal to edit it.
Start several plans at once with a launch set: the Brain lists the plans in the order to build them, each with its model and strength, and you press Start all (or Start first only). They build one after another through the project's queue, each after the one before it has finished every phase. Pressing Start twice never starts a plan twice. If one fails, the ones before it stay as they are and the rest wait; press Start again to carry on. Cancel stops only what has not started. The Brain never starts a build itself.
Directions ("every export is CSV first") and things you decided against reach every plan, phase and sort that follows. When something new resembles a decision you made, it is still recorded, with the earlier decision, its date and reason beside it so you can choose again — a match never silently drops a request. What a chat records can be undone for a minute, unless someone has changed it since.
The Brain remembers facts (with their evidence and where they hold), directions and decisions, and keeps notes. Everything it writes is its reading until you confirm it. It can write plans that have not shipped yet, but never a plan that is live, done, archived, or in a build that is queued or running. It cannot change your code, start a build, or confirm, archive or delete anything on your behalf.
Paste anything
The box at the top takes anything: a thought, a meeting transcript, a client email, a WhatsApp message. Pick where it came from (Thought, Meeting, Client, Email, Message) and, if you like, who or where ("Weekly call with Daniel"). Up to 150,000 characters at a time; a longer paste asks you to split it. You can also drop a .txt or .md file onto the box.
Your words are kept exactly as you pasted them, in the project at .nocturnal/drops/. They are never rewritten, and every item sorted out of them keeps the sentences it came from.
Paste the same text twice within a day and Nocturnal says so instead of sorting it again.
Sorting
Within a minute of pasting, a read-only Claude chat sorts the paste. It splits it into separate items — a feature, an improvement, a fix, a question, a decision, a note — places each on the product map, and for each one records:
- Whether it exists. Built (with the files, in this checkout), Built · live (only when a change or release record says so), Partly built, Not built ("I could not find … in the files I read" — never "does not exist"), or Could not tell.
- What would need adding.
- How big it is, as a size for an agent, never a time: tiny (one small change in one file), small (one agent phase), medium (a plan of two to four phases), large (several phases across areas, or a design or decision first).
- One sentence a client could read.
Pastes made within a minute of each other sort together as one chat, so ten quick lines are one run, not ten. Sort now skips the wait. Every sort is an ordinary chat you can open from its card, with its model and its cost on the Usage page — Included when it runs on your Claude plan, Cash when it uses an API key. The line under the box says which, before anything runs.
If a sort is stopped or fails, its card says why, with Retry and Add by hand (the original text beside an empty form, so nothing is lost). A sort that recorded nothing counts as failed, never as done.
Sorting never starts work. It cannot change the roadmap, create plans or start a build.
Turn automatic sorting off in Settings → Agents → Sort pasted ideas automatically; then a paste waits for you to press Sort. The same place picks the model. Sorting runs on Claude only.
Needs you
Sorted items wait in four piles: Looks already built, New, Questions, Could not tell (and agent Suggestions, when any are pending). Open an item to see what was said first, then what Nocturnal found, with links to the files or changes it cites — and whose reading it is.
From each item:
- Plan opens an unsent plan chat in Plans, already holding the original words and what Nocturnal found. Add your own context and send.
- Build opens an unsent agent chat in Building, the same way.
- To roadmap, Keep (for later) and Move (to another place on the map).
- Scrap asks why ("Already built" is filled in when it looks built). The item is kept with the reason, so the decision is not lost.
Select several and Plan together writes one plan for all of them.
When a chat started from Plan writes its plan, or a chat started from Build starts building, the item is marked with it. Once that chat's changes appear in the Changes log, the item shows where they have reached: built and waiting to be tested, or live.
The product map
The map is your product as areas ("Clients", "Settings → Billing") and, under each, what a user can do there ("Add a client"). Things that cut across the product, such as sign-in, live under Across the product. Each place has a dot: green built, amber partly built, blue not built, grey unknown — with a ring once you have confirmed it.
Map the product reads the project once (its screens, commands, README, plans and the Changes log) and writes a first map. It only runs when you press it; Rescan updates what it found and never renames or changes anything you confirmed, and never removes anything.
Click a place to open it beside the map: its standing and evidence, the items on it, plans, roadmap items, and the changes recorded there. "Built" stays Claude's reading until you press Confirm, or Correct it. An agent cannot confirm.
Drag an item onto a place to move it. Filter by Not built, Built, Needs decision or Unsorted, or search. On a phone the map is an outline.
Talking about the product
Discuss opens a product chat. It can read and change the map ("mark Add a client as built", "where does export fit?"), sort text you paste into the chat, and manage the Inbox, roadmap and open questions the way a project chat does. Roadmap changes still need your approval, and it cannot start a build. The pane beside the chat follows what it changed.
What syncs
The drop records, the map, and the items with short excerpts sync with your account, so the hosted dashboard and the phone show them. The pasted text itself stays on your computer; opening the original from elsewhere asks that computer for it. See What syncs to your account.