How to Set Up ChatGPT Projects for Recurring Work: Files, Instructions, and Memory

You ask ChatGPT for a weekly update every Monday, yet you keep pasting the same reference notes, restating the same style instructions, and correcting facts that were approved last week. Could a ChatGPT Project make this repeatable without creating a false sense that the AI remembers everything?

ChatGPT Projects organize related conversations, files, and project-specific instructions. They can make recurring work easier to manage, but they are not a verified database, an automatic publication system, or a substitute for human sign-off. This guide walks through a practical setup for a fictional community workshop preparing weekly public updates.

All example files and cases below are fictional. Official OpenAI documentation was reviewed on October 26, 2026. This guide does not claim that a ChatGPT account or interface was personally tested.

Choose a project only when context will be reused

A project is useful when several related chats depend on the same instructions and reference material. Weekly reports, recurring program updates, and ongoing research all fit this pattern. A single quick question usually does not need a dedicated project.

For our example, staff need to draft a short weekly program notice. Sources include an approved activity description, a current decision log, and a publication template. Budget, future event approvals, and private participant details remain outside the editorial workflow.

Before setup, state three boundaries: what the project can draft, what sources it may rely on, and what still requires human approval.

• In scope: draft a weekly update, summarize approved activities, identify unanswered questions.
• Out of scope: silently change event dates, make public announcements, contact participants, or publish content.
• Approval gate: a coordinator checks each date, location, contact detail, and link.

These are written editorial safeguards. They do not create application-level access controls by themselves.

Step 1: Create the workspace

Open ChatGPT and select New project in the sidebar, as described in OpenAI’s Projects help page. Name it something specific, such as Community Workshop — Weekly Updates, and choose an icon and color if desired.

Avoid combining unrelated work merely to minimize the number of projects. A project containing public event copy, confidential HR cases, and private client files is harder to govern than one with a clear purpose.

The product’s documentation says project availability is subject to plan and workspace settings. If the feature is absent, check your signed-in account and any organization restrictions rather than assuming that every interface is identical.

Step 2: Add trusted sources, not every old draft

According to OpenAI, project references may include PDFs, spreadsheets, documents, images, and pasted text. Project sources can be previewed or removed. The number of allowed files varies by subscription, so verify the current plan limits before committing to a large archive.

Start with a small, dated source set:

1. approved-program-description.txt — the current public program facts.
2. weekly-update-template.txt — approved structure and target length.
3. decision-log.txt — confirmed changes, owners, dates, and open questions.
4. source-register.txt — which version is authoritative and who maintains it.

Prefer approved and current files over lengthy drafts. If two sources conflict, preserve the disagreement and request review.

OpenAI also documents saving an AI response through its message menu using Add to project sources. Save a reviewed template or approved summary, not an unchecked answer. Otherwise an old AI mistake may be reused as if it were an authoritative record.

For supported accounts and connections, private projects can add Google Drive file or folder links and Slack channel links using Add sources. The official help explicitly says the Google Drive app does not provide advance sync when added in a project. Do not promise continuously synchronized content.

Step 3: Make project instructions operational

Open the project’s ••• menu, choose Project settings, and add project instructions. OpenAI says these instructions apply within the project and override global custom instructions there.

Use a short instruction that covers role, source priority, unknowns, output format, and approval:

Role: Editorial assistant for weekly community workshop updates.

Use sources in this priority:
1. Current approved program description
2. Most recent approved decision log
3. Weekly update template

Write a concise public draft with accurate dates and names.
Never turn a proposal into an approved event.
If a fact is missing or sources conflict, write "Needs confirmation."
Include a list of source filenames used and an approval checklist.
Do not email, publish, or modify external records.
A human coordinator must approve each final version.

The hierarchy makes it easier to spot stale information. The unknown-data rule discourages guessing. The approval instruction defines a human checkpoint; it is not proof that a separate publishing system will be blocked.

Revisit instructions after a workflow change. Prefer one clean current set to a growing stack of contradictory exceptions.

Step 4: Decide what memory may reference

OpenAI documents two project memory modes: Default memory and Project-only memory. In eligible accounts, open ••• > Project settings > Memory, choose a mode, and save. The setting change may require several hours to take effect.

With project-only memory, chats may use conversations inside the project, but they do not reference conversations outside it, and outside chats do not reference its conversations. Prior personal saved memories are not referenced within the project-only boundary.

Default memory behavior depends on plan and settings. Some accounts can draw on wider conversation context, while Enterprise and Edu have additional containment rules. Do not assume that default mode means the same thing for every user.

For a workshop with a narrow editorial mandate, project-only memory may be preferable. However, context isolation is not the same as confidentiality among project collaborators. If you share a project, other members can see shared project material according to their permissions.

OpenAI’s general Memory guidance also cautions that memory does not preserve every detail. Maintain exact approvals, amounts, and dates in source files, not in an assumption that a prior conversation will always be recalled.

Step 5: Use a repeatable weekly chat procedure

Start a new dated project conversation for each reporting period. This is an editorial naming convention, not a required ChatGPT setting.

A standard weekly request can look like this:

Draft this week's community workshop update.
Use the approved program description dated October 20
and the decision log dated October 26.

Apply the project's writing rules.
Identify the exact source filenames used.
Separate confirmed facts from "Needs confirmation."
Return a draft and a human review checklist.
Do not publish or contact anyone.

The four practical checkpoints are: confirm source versions, create the draft, reconcile important claims against files, and obtain human sign-off. Record the approved result outside the chat if the organization’s recordkeeping process requires it.

Projects collect context but do not automatically schedule or publish a recurring job merely because an instruction says “every Monday.” Scheduling and external delivery are separate supported workflows requiring their own controls.

Step 6: Move conversations selectively

OpenAI documents moving eligible prior chats into a project by dragging them into it or selecting Move to project from the chat menu. Moved chats inherit the project’s instructions and file context. Conversations created with a GPT cannot be moved into a project using that mechanism.

Move only relevant conversations. An old thread containing tentative plans may be harmful context if treated like a final decision. If a chat cannot be moved, begin a new project conversation with a reviewed handoff summarizing the facts that still matter.

Avoid importing personal or confidential discussions into a project you intend to share. Consider permissions before making past conversation context available to collaborators.

Step 7: Share only when the access model fits

OpenAI documents sharing projects through a Share control, with invitation or link options depending on plan and workspace policy. The owner can distinguish chat access, which lets members see and interact with project content, from edit access, which also allows modifying sources and instructions and inviting others.

Shared projects automatically use project-only memory and cannot be switched to default. This prevents a shared project from reading an individual’s outside personal memories. It does not hide material uploaded into that shared project from authorized members.

Use controlled invitations for internal content. Where “Anyone with a link” sharing is allowed, a signed-in user holding that link may be able to join. Review active members, visibility, and edit permissions before adding organizational files.

Deleting a project is not a harmless cleanup operation. OpenAI states that deleting it removes project chats, instructions, and project-only files, while separately saved Library material follows separate rules. Keep important records in the organization’s authoritative storage.

Test three situations before trusting the setup

A short review exercise reveals weaknesses more efficiently than asking the project whether it is reliable. The following cases use invented inputs; they were logically checked for this article but were not executed in ChatGPT.

Case A: known detail. A dated approved source says a workshop begins at 10:00 a.m. A passing draft preserves 10:00 a.m. and names the source.

Case B: missing detail. No source contains the next registration deadline. A passing draft labels the deadline “Needs confirmation” instead of inventing one.

Case C: conflict. An old draft says an event is Friday and a later approved decision log says Saturday. A passing draft recognizes the discrepancy, identifies the approved source, and does not silently cite the old draft as current.

Run these fixtures before relying on a real workflow, and again whenever sources or instructions change. The expected outcomes are editorial acceptance criteria, not a validated model-performance score.

Access limits and data controls

Projects are included with qualifying ChatGPT plans, but the number of files, collaborators, and available tools varies. Confirm current entitlements through the official Projects help page and your own account.

For personal accounts, OpenAI’s Data Controls guidance describes Settings > Data controls > Improve the model for everyone. Opting out of eligible model-training use does not delete stored chat history. Memory preferences, project sharing, and training controls are different settings; changing one does not automatically govern the others.

Organizations should follow their established confidentiality, retention, and access rules before uploading internal documents. A project is useful for organizing permitted work, not a shortcut around those obligations.

Final operational checklist

• Give the project one recurring purpose and an owner.
• Add approved, dated sources and document their priority.
• Write concise project instructions for unknown facts and review.
• Select memory boundaries with your actual account settings in mind.
• Use dated chats and a consistent weekly request.
• Test known, missing, and conflicting information.
• Review membership and permissions before sharing.
• Require human sign-off before publication or external action.

The goal is not to make ChatGPT appear to remember perfectly. It is to create a repeatable, inspectable process where sources, instructions, and approvals stay distinct.

Official sources and editorial scope

• OpenAI: Projects in ChatGPT — project creation, files, instructions, memory, chat moves, sharing, and limits.
• OpenAI: Memory in ChatGPT — personalization behavior and limitations.
• OpenAI: Data controls in ChatGPT — training preferences and history settings.

Documentation reviewed October 26, 2026. The workshop, filenames, prompts, and three test cases are invented examples. No project was created or tested in a user account, and no screenshot, vendor comparison, or live publishing result is claimed.