Claude gets worse the longer you run a single chat, and most people have no idea why. It is not your prompting. It is the context window, Claude's fixed working memory. Everything you connect and everything you type has to fit inside it, and once it fills up, Claude has less room to think. Here are the three fixes. All three take about two minutes.
Claude has a fixed working memory, called a context window. Everything you connect, and everything you type, has to fit inside that one fixed space. The longer a chat runs, the more of it is already spoken for.
As it fills, Claude starts compressing the earlier parts of the conversation to make room, so the detail you gave at the start goes fuzzy. The model has not got worse. It has run out of room, and the careful context from message one is the first thing to blur.
Claude's standard working memory, measured in tokens. A token is roughly a short word, so think of it as a long book's worth of room to think in.
Switch on too many connectors and tools and the space free for your actual work can collapse to around a third. The window is the same size. The clutter ate it.
The three fixes below. Turn off what you are not using, split your chats by topic, and save your context once. None of it is about prompting harder.
Picture a single whiteboard that never gets wiped mid-task. Every connector you switch on writes something in the corner before you start. Every message you send adds another few lines.
When the board is mostly full, Claude rubs out the oldest notes to keep writing, and those old notes were your setup. The fix is not a cleverer prompt. It is keeping the board clear enough to think on.
The model has a fixed amount of room. Spend it on your problem, not on clutter you forgot was switched on.
Anyone who uses Claude for real work and has felt it drift over a long session. The three fixes below are universal. The worked example runs on property, because that is what I use Claude for, but the habit transfers to anything.
Your email, your calendar, your drive, whatever you have linked. Each connector reserves a chunk of that working memory the moment it is switched on, before you have asked a single thing.
Most people leave everything on all the time. The result is a context window that is already part full before the real work starts, on every chat.
Go into your connectors and switch off the ones you do not need for the task in front of you. You are not deleting anything. You are clearing the desk for this piece of work, and you can turn them back on in seconds when the next task needs them.
When I am pulling comparables for a property deal, I keep only the connectors I actually need switched on. The answers stay noticeably sharper.
A connector Claude is not using is still taking up room it could be using to reason. Fewer connectors active means more of the window is free for your actual question. Keep your full set linked, but switch on only the handful this task needs, not everything you have ever connected.
Most people run everything through a single chat. Every task, every question, every unrelated thing, all in the same place. That chat fills up, and Claude starts mixing things that should never have met.
Use a separate Project for each thing you are working on, and start a fresh chat when you switch. Each Project is a clean window, so the context stays about one thing instead of a year of everything.
Open Projects and make one per topic you keep coming back to. When you move from one to another, start a new chat inside the right Project rather than carrying on in the old one.
This also locks in the first fix. Each Project only needs the few connectors that topic uses, so you are matching tools to the work without thinking about it twice.
I keep one Project per property deal or area I am researching, so Claude never blends two of them. The Dubai South numbers stay with Dubai South.
A fresh chat starts with a near-empty window. A six-month-old mega-chat starts almost full, and Claude is reading past noise it should have left behind. Smaller, single-topic chats keep the working memory pointed at the thing you actually asked about.
If you re-explain who you are and what you are working on every time you open a chat, you are spending working memory on the same paragraph again and again. And Claude still defaults to generic advice until you do.
Put that context into the Project instructions one time. Now every chat inside that Project already knows your situation before you ask a thing, without you spending a word on it.
Standing instructions for this Project. Apply them to every chat here, without me repeating them. Who I am: [your role and what you are working on, in a line]. Assume I know the basics of this area, so skip beginner explanations unless I ask. Play this role: act as a [the expert you want, e.g. seasoned analyst, sharp editor, sceptical adviser] who has done this many times. Bring that judgement, not neutral information. What this Project is for: [the one outcome everything here points at]. Always assume these facts: [the specifics that change your answer: key numbers, constraints, preferences, anything off-limits]. How to answer me: - Lead with your recommendation, then the reasoning, then the main risk or trade-off, then what I should do next. - Tight prose and short paragraphs. Bullets only for real lists or steps. Keep first answers short; I will ask you to go deeper. - Plain English. No filler, no flattery, no restating my question back to me. Never guess: if you are not certain, say so. Do not invent facts, numbers, quotes, or sources. Separate what you know from what you are inferring, and tell me how sure you are on anything I might act on. Challenge me: if an assumption of mine is weak, say so before you answer. Give me the strongest counter-argument, not just agreement. Ask first: if something you would need is missing, ask me one to three sharp questions before answering instead of guessing.
Project instructions load once and apply to every chat in the Project. You pay the memory cost a single time instead of on every conversation, and you never get the generic answer that comes from Claude not knowing your situation yet.
Message me directly if you want to talk through how to set up your Projects and connectors for the work you actually do. If property is part of your plan, I can also walk you through the opportunities I am tracking right now.
Message me on WhatsAppNone of this is about prompting harder. It is about protecting the working memory the model thinks in, so the answer on message fifty is as clear as the answer on message one.
Each one takes about two minutes, and they compound. A clean window, a single topic, and a standing brief mean Claude spends its memory on your problem instead of on clutter, crossed wires, and the same paragraph retyped.
It works for anything you use Claude for. I run mine on property deals, where the difference between a sharp answer and a vague one is the difference between a good decision and an expensive one.
Every prompt in this 30-day series is one capability. FourthspaceOS bundles all of them into a single product: underwriting, comps, market research, deal sourcing, portfolio tracking, and investor reporting. The product runs natively on the same Anthropic agents and plugins you are learning to use this month.
Waitlist members get founding pricing, early access, and one new prompt delivered each day for the next 30 days.
Founding pricing locks in for waitlist members.
One email a day for 30 days. No spam. Unsubscribe whenever.
The three fixes you just read are Day 22 of 30. The full series, daily builds, and live property workflows are on Instagram. Follow for the rest of the run.
If you found this page through the reel, the feed is where the next one lands. One prompt, one fix, one workflow at a time, all built for property investors.
@fourthspaceOS · new prompt every weekday
Follow on Instagram ↗Free. Daily. Unfollow whenever.