What is actually going wrong
You are running everything in one long chat. It works well for the first hour. Then it starts suggesting things it suggested earlier, and mistakes you corrected an hour ago come back as if the conversation never happened.
Nothing is broken. Every model has a limit on how much of the conversation it can hold in front of it at once. As the chat gets longer, the earliest part of it stops being read in full. Your correction was in the earliest part.
So the fix is never a longer chat. The fix is getting the important things out of the conversation and into somewhere permanent.
Fix 1: if it is a repeatable task, use a Project
Anything you do more than once belongs in a Project rather than a chat. Put the relevant files in, and give it the correct instructions on how to do the job properly, so it stops making the mistake in the first place.
The difference is where the correction lives. In a chat, a correction is a message that scrolls away. In a Project, a correction is a line in a file that gets read at the start of every session, forever.
When you correct Claude on something and the correction matters, do not just say it. Ask it to add the rule to the CLAUDE.md. Takes five seconds and you never have to say it again.
Day 3 is the full Cowork Projects walkthrough, including the five files that go in one and three Projects you can copy.
Fix 2: if it is one big job, use Claude Code and /compact
Some jobs are not repeatable. They are one large piece of work that will make a real difference to the business, and they run long enough that the chat fills up part way through.
For those, get a paid Claude plan. Even Pro will do. What you are buying access to is Claude Code, and specifically the /compact command.
What /compact does
Type /compact and Claude condenses the conversation so far into a summary and carries on from that, instead of hauling the entire history along behind it. The important decisions survive. The long back and forth it took to reach them does not.
Less to read means less to lose track of, which means fewer of those repeated mistakes.
/compact is a Claude Code command. Cowork manages its own context in the background with no manual trigger, and Chat has neither.
If a chat has got so long that you are wishing you could compact it, that is usually the signal it should have been run in Cowork or Code in the first place.
The blockers file, which is the bit that actually saves you
Compacting keeps the thread going, but a summary is still a summary. Some of the detail goes.
So on any long job, get Claude to write the mistakes down as they happen. Ask for a markdown file that logs every blocker and every correction while you work. Then the record lives in a file rather than in a conversation, and it survives a compact, a crash, or you closing the laptop and coming back on Thursday.
As we work, keep a file called blockers.md in this folder. Every time we hit one of these, add a line to it: - a mistake you made that I corrected - a dead end we went down, and why - a decision I made that changes how the rest of this job is done Each line: what happened, then the rule to follow instead. Read blockers.md back at the start of every session, and again straight after any compact, before you do anything else.
Two weeks into a big build, that file is the most valuable thing in the folder. It is the only place the hard-won corrections are written down.
Where to put a file, depending on where you are working
The other half of the same problem: feeding Claude the right thing in the right place.
| Where you are | How you give it a file | Use it when |
|---|---|---|
| Chat | Drag and drop, or the attach button | A one-off. Something it needs to see once and will never need again. |
| Chat, in a Project | Add it to the Project's files | Something every conversation in that Project should know. Your voice, your templates, your rules. |
| Cowork | Put it in the folder you trusted | Anything the job produces or works from. It reads the folder and writes back into it. |
| Claude Code | Point at it by name, or drop it in the folder | Anything in the project. It can go and find files itself once it knows the folder. |
Ask yourself whether you will need to give it this file again next week. If yes, it belongs in the Project or the folder, not attached to a message.
When to start a fresh chat
Four signals. Any one of them and you are better off starting again:
- It repeats a suggestion it already made.
- It brings back a mistake you corrected earlier.
- You have changed subject and the old thread is only weighing it down.
- You are on a different job now, even a related one.
Starting fresh costs you nothing when the context lives in a Project or a folder. That is the real point of everything on this page: the value should be in a file, not in the chat you are frightened to close.
Book a free 30-minute AI assessment
Thirty minutes, no pitch. We work out which system would make the biggest difference in your business, and you leave with the answer whether you ever hire me or not.
Book the assessment