GPTMap

ChatGPT Projects Complete Guide: File RAG, Custom Instructions, and Team Collaboration

ChatGPT Projects is the workspace feature for Plus/Pro users: file uploads, project-level Instructions, cross-conversation context, and workspace sharing. A copyable SOP for setting up a project.

TL;DR
ChatGPT Projects is the workspace feature for Plus / Pro users that lets multiple conversations share the same context: file uploads (auto-RAG), project-level Instructions, cross-conversation Memory, workspace sharing. This article gives a copyable setup SOP covering when to use Projects, file naming, Instructions, team collaboration, and the relationship with Memory / Custom GPTs.
ChatGPT Projects is the workspace feature for ChatGPT Plus / Pro / Team / Enterprise users, organizing multiple conversations that share the same context: project-level Instructions, uploaded files (auto-RAG), and project-level Memory.

How to

  1. Decide if it's worth opening a Project

    Triggers: same project for ≥3 conversations / same reference doc re-used / cross-conversation context needed - any one is enough. Otherwise stay in regular chats.

  2. Name + one-line description

    Visible in the Projects list. Use a concrete name ('GPTMap batch-4 writing') rather than 'My project 1'.

  3. Upload files with descriptive names

    Drop in the references you need: long PDFs (reports / contracts), code, CSV. Name descriptively ('2026-product-pricing.pdf').

  4. Write project-level Instructions

    More specific than global Custom Instructions: project background + role + behavior rules + glossary. 100-500 characters is the sweet spot.

  5. Enable Memory (project-level)

    Settings → Memory → enable project-level Memory. Memory is shared across all conversations in the Project.

  6. Test the first conversation

    Open the first conversation in the Project: test typical questions (confirm Knowledge citations + Memory accumulation); tune Instructions; start using.

ChatGPT Projects is the workspace feature for Plus / Pro users - organizing conversations that share the same context.

1. When to use Projects

Projects solves "multi-conversation shared context." Three typical scenarios:

  • Reading the same book three times: working through a long PDF / paper across multiple sessions
  • Same project for weeks: a writing project, data analysis, product research
  • Same reference doc re-used: team-shared product specs, customer case library

Scenarios that don't fit Projects:

  • One-shot Q&A ("help me write an email") - use a regular chat
  • Want to share with many external users - use Custom GPTs + GPT Store
  • Multi-modal creation (image, audio, video) - use specialized tools

2. Create a Project

ChatGPT left sidebar → Projects → New Project:

  1. Name + color + icon: identifies it in the project list ("GPTMap batch-4 writing" beats "My project 1")
  2. Upload files: drop in references
  3. Write project Instructions: project-level prompt
  4. Enable Memory (optional): project-level Memory

3. Uploaded files: auto RAG

Uploaded files are auto-indexed. When answering, the model:

  • pulls relevant snippets
  • shows the filename as the citation source
  • is great for "ask precise questions about project docs"

Important boundary:

  • The model cannot reproduce long PDFs verbatim - it does RAG retrieval
  • Key facts (company background / product specs / glossary) should go directly into project Instructions as the authoritative source
  • Name files descriptively: 2026-product-pricing.pdf is 10× better than pricing.pdf

4. Project Instructions: project-level prompt

Project Instructions override global Custom Instructions. Recommended structure:

# Project background
This is [project name], goal [one sentence].

# Role
You are [role] in this project, expert in [specialty].

# Behavior rules
- Use [terminology]
- When citing files, mark [filename]
- Answer in [format]

# Glossary
- Term A = [definition]
- Term B = [definition]

100-500 characters is the sweet spot. A glossary is especially useful - it stops the model "guessing" domain terms every conversation.

5. Memory: shared across conversations

Project-level Memory is shared across all conversations inside a Project. Example:

  • Conversation #1 inside the Project: "I'm working on GPTMap batch-4 writing"
  • Conversations #2, #3, ...: the model still remembers

Memory survives project end - clear it manually or start a new Project.

Caveats:

  • When Memory accumulates 200+ entries, the model can get information-overloaded
  • Periodically review, delete stale / wrong entries
  • Project Memory doesn't cross projects - explicitly copy or re-set when needed

6. Team collaboration

TierProjects sharing
Plus / ProOnly yourself
TeamAll workspace members
EnterpriseWorkspace members + configurable by dept / project

Team-shared Projects caveats:

  • File privacy: uploaded files are visible across the workspace - handle sensitive data carefully
  • Shared Instructions: all members see the same prompt - team-prompt accumulation has value
  • Shared Memory: every member's conversations contribute - can produce inconsistencies

7. Projects vs Custom GPTs

DimensionProjectsCustom GPTs
Primary useMy own multi-conversation workDistribute to others
FilesUploaded to ProjectKnowledge upload
InstructionsProject-levelGPT-level
MemoryProject-levelNone (unless Code Interpreter etc.)
ActionsNoneStrong (external API)
Team sharingTeam / EnterpriseOnly via GPT Store publish
DistributionWorkspace onlyGPT Store public

Common combo: Custom GPT provides the conversational surface (Actions / Knowledge); Projects manages the team's multi-turn follow-up (context, Memory).

8. Practical SOP

Scenario: 4-week product research project

Week 0: Create Project "Q3 product research - competitive analysis"; upload competitor PDFs + annual financials; write project Instructions (glossary + role); enable Memory.

Week 1-4: Open a new conversation each week - "Q3 research weekly progress" - referencing prior Memory + files.

End of week 4: Export key conclusions to Notion / Confluence; keep Memory for future reference.

9. Common errors and troubleshooting

  • Files uploaded but model ignores → descriptive names / model can't find snippets; rename or put key info directly into Instructions
  • Memory accumulated too much → periodically review; delete stale / wrong entries
  • Cross-Project context mixed up → Memory doesn't cross projects; use explicit naming to keep projects distinct
  • Team-shared file privacy leaks → don't put sensitive data in Projects; use Enterprise workspace permission controls
  • Instructions written like prose → behavior becomes unpredictable; switch to structured sections

10. What's Next

  • ChatGPT Memory: The Complete User Manual - Memory details
  • Custom GPTs Complete Guide - full Custom GPTs path
  • ChatGPT Complete Guide (2026): From Beginner to Expert

Update log

  • 2026-08-08: Initial publish

Key points

  • Projects solves the 'multi-conversation shared context' problem; single-shot Q&A stays in regular chats
  • Uploaded files are auto-RAG'd but the model cannot precisely reproduce long PDFs - put key facts directly into project Instructions
  • Project-level Instructions override global Custom Instructions
  • Memory is shared across all conversations in a Project; survives project end (this is a feature, not a bug)
  • Team / Enterprise users can share Projects inside the workspace; Plus / Pro users see their own only
  • Projects is not a replacement for Custom GPTs - Projects for 'my own multi-conversation work', Custom GPTs for 'distribute to others'

Frequently asked questions

Regular ChatGPT is single-shot: each new chat starts clean. Projects organize multiple conversations into a workspace sharing the same uploaded files, project Instructions, and project-level Memory. Ideal for 'reading the same book three times' or 'working on the same project for weeks'; not for single Q&A.

Official references

Related articles

Subscribe to GPTMap Weekly

One email every Monday: curated OpenAI updates, deep dives, and best practices. No ads, unsubscribe anytime.

GPTMap EditorialPublished 2026-08-08 5 min read
Test environment (EEAT)
Last tested: 2026-08-08
Model used: gpt-5.6