You know the first loss. The day you stop. Life gets in the way, energy drops, the project goes quiet. That one you can see coming.
The second loss is quieter. It happens when you come back. You open a new conversation and Claude doesn't know who you are, doesn't know what you built, doesn't know what you decided three weeks ago. So you start explaining. And somewhere in the explaining, you feel the thing you built getting smaller, not because it's gone, but because you can't find your way back into it. It feels like a lost connection - one that has to be re-established.
Most long projects don't fail at the hard part. They fail at the re-entry.
The last module handed you the loop - canon, pack, ignition. That system carries the work between threads. This module builds the room the system lives in.
Claude Projects is a container, a dedicated workspace where your files, your instructions, and your conversation history all persist between sessions. You open it and Claude already knows where you are. You pick up instead of starting over.
Open a Project and the whole room is laid out on one screen. The project’s name sits at the top with the one-line description you gave it. The composer waits in the middle - every new thread starts there. Beneath it, your conversations: the ones you pinned, then the recent ones, each stamped with its age.
Down the right side, three panels carry the continuity. Instructions, with a pencil beside it - that opens the brief you write once, and we’re about to write it. Memory - the running notes the platform keeps about the project as you work, visible only to you. You can’t type into the panel, but you can prune it right there. And from any thread you can tell the machine what to remember, alter, or drop - it recalibrates almost on the spot.
And the third panel may be the most important of them all.
Context is the quiet nexus of the room. Every thread you open here reads against it before a single word comes back - not storage but calibration, running in real time. The shelf holds your files, each one a card showing its name and size, with a meter underneath reading how much of the room’s capacity you’ve used. But what it’s really holding is alignment: your voice, your decisions, your world, kept within arm’s reach of the machine at all times.
You picked the machine up as an instrument of your craft. The shelf is the machine’s own toolkit: the reference it reaches for the way you reach for it. Equip your tool, and everything it hands back arrives pre-aligned.
Labels and placements will drift as the platform evolves. The architecture won’t: one brief, one shelf, one running memory, many threads.
But the container is only as good as what you put in it. This module is about building it right, so that coming back feels like walking into a room that's been waiting for you, not a room you have to rebuild from memory.
Every Project has Custom Instructions - the panel on the right, behind the pencil. This is where you establish the creative relationship before the work begins, your voice, your standards, what you need Claude to do and not do. Write it once and it applies to every conversation in the Project forever. No re-explaining. No re-orienting. Claude walks in already briefed.
The instructions aren't a description of the project. They're a description of the partnership. How you write. What honest feedback looks like for you. What role Claude plays. What to show first and when to explain. Write it like you're briefing a collaborator who's about to spend the next year with you.
One thing the panel will never do is answer. Custom Instructions is a settings field, not a conversation: you file the brief, and every thread after simply behaves like it read it - because it did.
The pro move before you file: run the brief past the machine itself, in any thread.
What came back is a working file. If it fits the way you work, lift it whole, swap in your own reader and your own rules, and file it.
The Context panel is your briefing book: the documents Claude can reference in any conversation without you pasting them in. Your canon goes in first - the living ledger from the last module now has a permanent address. Your voice document. A sample that captures the tone you're building toward. Character or subject profiles. Any structural document, outline, timeline, chapter map.
Start lean. Four-to-five files is enough to orient a first session completely. Upload what shows, not just what tells. A well-chosen sample scene tells Claude more about your voice than a ten-page description of it. And the capacity meter is not a challenge: a shelf reading single digits that orients a session completely is the goal, not a shortfall.
One instrument stays off the shelf: the LaunchPack™. The pack rides in the first message of every new thread, pasted, never filed. Inside a Project it runs short - your position and today’s job - because the room already carries the rest. Outside one, it carries the full load. Either way it gets cut fresh from the canon every time, because a pack is momentum, and momentum doesn’t keep.
And underneath that rule sits the distinction that makes the whole panel click - the one most operators never learn. The shelf is not storage; it is the room's standing knowledge. Everything filed there gets read into the machine's awareness in every single session, which makes it the right home for law and the wrong home for cargo. Files the machine must work on - the manuscript, the assets, the build - ride in the message instead, where they land in its hands rather than its head. The shelf carries what must be known at all times. The thread carries what must be touched right now. Learn that split and the bandwidth takes care of itself: law stays light, cargo stays out of the way, and nothing pays rent it doesn't owe.
Inside a Project, every conversation thread is saved. Name them when you open them - not "New Chat" but "Chapter 6 - the confrontation" or "Voice, testing third person." They land in that pinned-and-recents list on the home screen, and three months from now you'll scan it and know exactly where everything lives.
Know the difference between those two lists, because they measure different things. Recents is automatic - the platform sorting by whatever you touched last. A record of activity, nothing more. Pinning is a decision: your declaration that a thread matters no matter when you last opened it. The busiest thread is rarely the load-bearing one. Pin what the project stands on - the master draft, the canon - and let Recents carry the churn. A pinned thread survives the scroll: three weeks untouched and still first in the room.
The word comes from construction. A load-bearing wall isn't the biggest one or the most beautiful one - it's the one holding the floors above it. Pull it, and the house comes down. In a creative operation the test is identical: if this thread, this file, this decision disappeared tonight, what else falls with it? The canon is load-bearing. Your voice document is load-bearing. The brainstorm where you riffed on chapter titles is drywall - useful, replaceable, never structural. Motion is not the measure. Weight is.
And don't be afraid of emojis in the title box. A flame on the active draft, a lock on the sealed chapter, a magnifier on the research thread - glyphs bring real value to cataloging, because a scanning eye lands on a symbol faster than it lands on a word.
Keep threads focused. When one starts circling across too many problems, start a new one. The old thread stays exactly where you left it. Focus is the feature.
When a thread runs weeks-long, Claude's recall starts to compress at the edges. Ask for a summary and you might get the last third, not the whole arc. The fix is an anchor, go back to the very first message, copy it, and ask Claude to build a complete table of contents with that message as the start and the current message as the end.
The bookends force full recall. This technique came out of the 46-day single-thread collaboration that became this guide. Without the anchor, half the work would have been unrecoverable.
Every three to five sessions, stop and ask Claude to account for where the project actually stands, not where you feel it stands. What's been decided. What's locked. What's still open. What's being circled without resolution. The third question is the one that does the real work, it names the thing you already know is there but haven't faced yet.
Projects require a paid plan. If you're not there yet, the same principles apply, you carry the context manually. Your canon does the heavy lifting here: voice brief, current state, decisions, one document. Paste the essentials at the start of each session, and close each session by filing to it. When you move to Projects, the canon walks in as your first upload.
Don't wait until you're deep in a project to set up the container. Build it now, while the work is fresh. Create a Project. Write the Custom Instructions as if you're briefing a collaborator who's about to spend the next year with you. Upload one sample that shows what the work sounds like.
Then ask Claude one question: what does it know about this project based on what's already there? The answer shows you what's clear and what's missing. That's your baseline. That's the room before anything goes wrong. The second loss is preventable, but only if you build the room before you need it.