How Do You Keep a Long Claude Draft Consistent Across Chats?
Keep a short decision log listing the claims made, terms defined, figures used and choices settled so far. Save it in your project's knowledge, and update it after each section. Anthropic's help page says context isn't shared across chats in a project unless it's in the knowledge base, so the log is what carries your decisions forward.
We're writing [document] in stages across several chats. The attached decision log is the record of what's been settled, so don't contradict it. Write the next section: [section name and goal]. After the section, give me an updated decision log in this exact format, as Markdown I can save: ## Claims made - [claim] (section N) ## Terms and definitions - [term]: [definition as used] ## Figures - [number, what it refers to, source] ## Decisions - [choice we settled and why] ## Open questions - [anything unresolved] If this new section would conflict with the log, tell me where before you write it. Keep the log under 400 words by removing anything that no longer affects later sections.
- Claims made
- Terms defined
- Figures used
- Choices settled
Try it now
- 1Click **Projects in the left panel, then + New Project**, and name it after your document.
- 2After your first section, ask for a decision log and save it as a .md or .txt file. In the project's knowledge panel, click + and upload it.
- 3Start a new chat inside the project, and paste the prompt for the next section.
- 4You'll know it worked when Claude flags a conflict with the log, or uses your exact defined term in a chat that has never seen the earlier sections.
Use it when
Books, course material, long reports, and any piece written over days or across several chats. It helps most when figures or defined terms have to match across many sections, or when two people are drafting different parts.
Skip it when
A piece you'll finish in one sitting and one chat. Claude already has the earlier sections in that conversation, so a log just adds work. It's also unnecessary if you're working in one doc and the whole text is short enough to attach to every new chat.
Why it works
A long piece usually spreads over several chats, and each new chat in a project starts without the others. Anything in project knowledge is available to all of them. A log is much shorter than the draft, so it's quick to check and cheap to keep around.
A single long chat has a different problem. On paid plans with code execution on, Anthropic says Claude summarizes earlier messages as a conversation nears the limit. Your full history is preserved and Claude can still reference it, but long conversations that trigger this use more of your limit. An explicit log means your key decisions don't depend on what a summary kept. Anthropic also says project content is cached and counts less when reused, though caches expire after inactivity.
Example
Say you're writing a 12-chapter employee handbook. In chapter 2, you decide "contractor" means anyone paid by invoice, and that the probation period is 90 days. Both go into the log. When you draft chapter 9 in a new chat, Claude can see those entries. If your notes for chapter 9 mention a three-month probation, it can flag the mismatch rather than quietly writing a different number.
Going further
Replace the file, rather than adding a new one each time. Rule of thumb, not a vendor figure: old versions left in the project can compete with the current one, especially since large projects on paid plans switch to retrieval. Anthropic's RAG page says Claude then searches for relevant pieces instead of loading everything, and it recommends clear filenames, grouped content and naming specific documents in your question. Call the file decision-log.md and mention it by name.
Don't treat project memory as the log. Anthropic says each project has its own memory, but a file you control is easier to check line by line, and your figures need checking.
Common mistakes
- Assuming chats in the same project can see each other. Anthropic's page says they can't, unless the content is in the knowledge base.
- Letting the log grow into a second draft. A log of 3,000 words is one nobody checks, so ask Claude to prune it.
- Keeping the log in chat only. The next chat can't see it, so it has to be saved in the project.
- Staying in one enormous chat to avoid all this. Anthropic says conversations that trigger summarizing use more of your limit.
Sources
- How can I create and manage projects?, Claude Help Center
- Context window on paid plans, Claude Help Center
- Usage limit best practices, Claude Help Center
- RAG for projects, Claude Help Center
Checked 6 November 2026. Products change often, so confirm details with the vendor.