AI Instructions vs Knowledge: What Goes Where?
Direct answer: Put directions for how the AI should work in instructions. Keep the documents, examples, conversations, and reasoning it should work from as retrievable knowledge. Connect the two with a small, specific request: which source to use, for which task, and what evidence to show.
Your working rules can be short without making your thinking smaller.
Perhaps your project instructions started as three helpful sentences. Now they contain your business history, meeting notes, a writing guide, and several reminders added after disappointing answers. Editing them has become another job.
You don't need to throw that material away. You need somewhere for it to live where you can bring the right part back.
Multiplist gives you a vault for chosen conversations and documents, with tools for saving, searching, and retrieving sources. That lets you keep the fuller thinking available while using a shorter instruction to direct the next task. It is useful when the same work moves between connected assistants. Saving and retrieval are deliberate steps, rather than a promise that everything you've ever said will appear automatically. Multiplist tool reference
Connect Multiplist to Claude. Sign in to the intended Multiplist account, review and authorize the connection, enable it where needed in Claude, and save the material you want to reuse. Then ask Claude to retrieve a known source before relying on it.
# Try the distinction on one real task
Imagine you're preparing a workshop invitation. You have an approved outline, a conversation about who the workshop serves, and a few invitations that sound like you.
Here is a useful division:
- Instruction: “Before drafting a workshop invitation, retrieve the approved outline. Use its audience and scope. Ask if the date or booking link is missing.”
- Knowledge: The actual outline, the original audience discussion, and the example invitations.
- This task's request: “Write a short invitation for the October workshop using the approved outline dated October 12.”
The instruction establishes a repeatable way of working. The sources contain the details and reasoning. The current request selects the job.
A sentence such as “the workshop lasts two hours” is a fact you may want the assistant to retrieve. “Confirm the workshop duration from the approved outline before writing” is a working rule. Keeping those roles clear makes a changed duration easier to handle.
This is a practical organizing method, not a requirement to reorganize your whole business.
# A five-step setup you can finish with one document
# 1. Choose the next task, not your entire archive
Start with something you actually repeat: a proposal, a weekly review, or a product explanation. Find the existing source you trust for that task.
If the useful thinking lives in a transcript, preserve that transcript. You do not have to replace it with a freshly generated summary to make it valuable.
# 2. Save the reference material with a recognizable title
Use a title that distinguishes the source from related drafts, and include a date or version when that matters. For example: “Workshop outline — approved October 12.”
Multiplist's source tools support saving text and retrieving a source by identifier or title. Keep the returned identifier if similarly named documents might be confused. Source capture and retrieval
Only save material you are entitled to share with the services involved. Review the Multiplist privacy policy and your assistant's account or organization rules before adding confidential work.
# 3. Write a small instruction that points to the evidence
Try:
“For workshop invitations, retrieve ‘Workshop outline — approved October 12’ from Multiplist before drafting. Identify the source you used. If it is unavailable or conflicts with my current request, tell me before assuming the missing details.”
A title in instructions is a pointer, not proof of access. The assistant still needs an available, authorized connector and a successful retrieval.
# 4. Give a task and inspect the result
Ask for one invitation. Check the audience, duration, and booking details against the source yourself.
If the assistant didn't retrieve the source, adding more rules may not help. First check the connection and the source title or identifier. Our guide to checking whether AI read your saved source walks through that separate verification step.
# 5. Maintain the source when the facts change
When the workshop changes, save the revised material and make clear which version is approved. Update the instruction's pointer if you used a version-specific title.
If you also uploaded copies into several native projects, treat them as separate copies until you have verified how they update. For conflicting old information, see how to correct outdated AI memory.
# Where this fits in Claude and ChatGPT
Feature details checked October 17, 2026. Interfaces and availability can change; follow the current vendor documentation for your account.
# Claude projects
Claude separates project instructions, which guide its behavior across project chats, from project knowledge, where you add relevant files and text. This is a good native home for a self-contained project. Team and Enterprise sharing can depend on administrator settings. Claude project documentation
You can keep the project's working directions there and choose whether its reference material belongs in project knowledge, Multiplist, or both. Use both only when the extra copy serves a purpose.
# ChatGPT projects
ChatGPT projects bring chats, reference files, and project instructions together. Project instructions apply within that project and override global custom instructions. File limits depend on the plan, and workspace controls still apply. ChatGPT project documentation
If everything you need stays in one project, that may already meet your needs. A portable vault becomes useful when you want to reuse selected source material in other connected tools without rebuilding the research each time. Check your actual connector availability before designing a workflow around it.
# Claude Code
For coding work, Claude Code's CLAUDE.md provides persistent project instructions. Anthropic recommends keeping these files concise and specific. Putting a large reference document into an imported instructions file still adds it to context; an import is not a way to make that context cost disappear. Claude Code memory documentation
Keep the instructions focused on how to work in the repository. Retrieve additional background when a task needs it. A local instructions file also doesn't, by itself, connect Claude Code to your Multiplist account.
# Keep authority separate from reference material
An old brainstorming transcript may contain “let's double the price” alongside a later decision not to. A quoted example may contain directions meant for somebody else.
Tell the assistant which document is approved and what role it should play. Ask it to surface conflicts instead of quietly choosing a convenient sentence. Retrieved material is evidence to interpret, not blanket permission to act.
The aim is a calmer setup: a few clear working rules, the original thinking preserved, and a way to bring the relevant source into the task at hand.