ChatGPT and Claude Projects: Set Up Reusable AI Context
Table of Contents
Return to the Practical AI Workflow Course
Create a project when several conversations share the same purpose. Put stable instructions and approved reference files in one place. Keep each conversation focused on one task. This lesson uses a fictional community workshop for practice without private data.
Learning outcome: you will create a project with two controlled sources and test its facts in a new chat.
Key Takeaways
- Project instructions set reusable behavior for related chats.
- Project sources hold reference material you want the assistant to consult.
- Conversation history is a record of one exchange, not a substitute for an approved source.
- A fresh-chat test reveals whether your project setup supplies enough context.
Before You Begin
Prerequisites: a ChatGPT or Claude account, plus the task packet from lesson 1 . Set aside 35 minutes. Difficulty is beginner. Use invented names, dates, and locations throughout this practice.
Create two small files on your computer. Open a plain-text editor, paste each block below into its own file, and save with the shown .md name. Markdown files are plain text. workshop-brief.md states the audience and style. approved-schedule.md states the fictional official schedule.
workshop-brief.md, revision 1
Audience: people registered for an introductory workshop.
Style: short sentences and a friendly, factual tone.
Do not invent prices, accessibility features, or parking details.
approved-schedule.md, revision 1
Source marker: WS-2026-11-12-R1
Event: Introduction to Digital Safety
Date: 2026-11-12
Start: 7 p.m.
Place: Community Room A
Set Up ChatGPT
- Open Projects in the ChatGPT sidebar and create a project named
Workshop Practice. - Add the two files as project sources.
- Open Project settings and add the instructions below.
- Inspect memory and sharing settings for your account and workspace before adding real material.
Use approved-schedule.md for event facts.
Use workshop-brief.md for audience and style.
If a requested fact is absent, ask for the source or mark it unknown.
Draft messages only. Do not send or publish them.
List factual changes after each revised draft.
ChatGPT Projects keep related chats, files, and instructions together. Project settings and available features vary by plan and workspace. Choose project-only memory when the creation flow offers it for a clean isolation test. Under default memory, a chat might draw on information outside the uploaded files. The unique source marker below helps you test the file without treating an answer from memory as file evidence. Use the official ChatGPT Projects guide to check current memory options.
Set Up Claude
- Open Projects in Claude and create a project with the same practice name.
- Add the two files to project knowledge.
- Set project instructions to the same five lines.
- Start a new chat inside the project to test the setup.
These steps describe classic Claude Projects. Anthropic is rolling out a separate Projects beta to select Pro and Max Claude Code users. In the beta, one project is one conversation with parallel cloud threads. Check What are Projects? and follow your account interface. Classic project instructions apply across chats, and project knowledge supplies shared reference material. Use the uploaded files as the controlled facts. Project memory behavior varies by account and settings, so do not treat another chat’s details as the approved schedule. Review Claude’s project setup guide for current controls.
| Project mode | Shared material to inspect | Fresh test |
|---|---|---|
| ChatGPT project | Uploaded files, instructions, selected memory mode | New project chat cites the schedule marker |
| Claude classic project | Project knowledge and instructions across chats | New project chat cites the schedule marker |
| Claude Projects beta | Sources and context in its project and cloud threads | New project thread cites the schedule marker |
A project is a context container, not a source approval. The marker shows whether a named file reached the chat. You still compare date, time, and room with the file yourself.
Test a Fresh Chat
Open a new ChatGPT project conversation or a new classic Claude project chat. In the Claude Projects beta, use a new thread in the project. Ask for a short announcement. Omit the schedule from your prompt so the project sources do the work.
Draft a 100-word email for the upcoming workshop.
Name the source file and its exact source marker for the time and place.
List any missing facts after the draft.
Do not send the message.
| Result | Meaning | Next action |
|---|---|---|
| Time, place, and marker match | The file is available to this chat | Confirm the facts against the file yourself |
| A fact is invented | Instructions or source use failed | Remove the fact and narrow the request |
| Source is absent | The reference did not reach this chat | Check project membership and file upload |
| Old revision appears | The project holds stale material | Replace the source and retest |
Do not infer reliability from one correct answer. Default project memory might supply a familiar time without reading the uploaded file. Require the exact WS-2026-11-12-R1 marker, then repeat the test with this conflicting draft:
Revise this line using the approved project schedule:
"The Introduction to Digital Safety class starts at 6 p.m. in Community Room A."
Name the factual change, source file, and source marker. Do not send the message.
Expected answer: the line uses 7 p.m. and names approved-schedule.md with marker WS-2026-11-12-R1. Record any failure you observe. A passing run needs no invented failure.
Worked test record: new chat, first prompt, time 7 p.m., place Community Room A, source approved-schedule.md, marker WS-2026-11-12-R1, pass. Conflict prompt, 6 p.m. corrected to 7 p.m., pass. This is an example record. Replace it with your own observed results.
Maintain the Project
Keep project instructions short. Put stable rules there. Put task-specific details in each prompt. Replace outdated files and identify their revisions. Separate unrelated work into another project so sources do not conflict.
Before sharing a project, inspect its files, chats, instructions, and access list. A project is a work container, not an approval system. A source file named “approved” still needs a real owner and revision record for consequential work.
Practice and Check
Your artifact: one project with two synthetic source files, five instruction lines, and a short test record. Capture the observed time, place, source name, marker, and any invented fact. Delete the practice project when finished if you do not need it.
Expected reasoning: a new chat uses project sources for stable facts, while your prompt defines the immediate output. You verify the result against the file. A missing source or wrong revision sends you back to project setup.
Check Your Understanding
Question: Does a new ChatGPT project chat guarantee isolation under every memory setting? Answer: No. Check project memory mode and test with a unique source marker.
Question: What should you do when the project repeats an older schedule? Answer: Inspect the attached source and memory scope, then replace stale source material.
Troubleshooting
- Missing project controls: check your current plan and workspace settings in the vendor guide.
- Conflicting answers: remove duplicate or stale source files, then start a fresh chat.
- Unexpected sharing: review project membership and remove sensitive sources before continuing.
- Repeated style drift: shorten instructions and provide one short example in the task prompt.
Next Steps
Continue with AGENTS.md Repository Context . The next lesson moves from a chat project to a local code repository.
Course navigation: Previous: Workflow Foundations , Course outline .
