ChatGPT Projects: your first context workflow (2026)

ChatGPT Projects is the space on https://chatgpt.com where chats, files and instructions live together for one body of work. It is not another loose thread: it is a container with rules and sources the model can use on every message inside the project. The official guide is at https://help.openai.com/en/articles/10169521-projects-in-chatgpt. This flow takes about 20 minutes and leaves you a reusable project for repeated work — a client, a report type, a research line that grows over weeks.

1. Steps 1

   Decide if you need a Project or a plain chat

   Before clicking New project, answer: will you repeat this work several times with the same rules and documents? Or will the thread grow over weeks with many files on the same line of work? For a one-off email or quick question, stay on the normal chat (or a saved prompt). For a weekly brief, client reports or research that accumulates sources, a Project is worth it. If the case is a fixed role you open in short sessions without a huge thread, read chatgpt-gpts-cuando first. Write the purpose in one line: «Monthly reports for Client X with glossary and direct tone.»

   What you should see: A one-line purpose in a notes app. A written decision: Project, plain chat or GPT.

2. Steps 2

   Create the project on chatgpt.com

   Go to https://chatgpt.com. In the sidebar, click New project (or open Projects in the left menu and create one). Give it a concrete name — not «Writing» but «Client X Reports» or «Product brief Q4». Optionally pick an icon and color to spot it quickly. The name tells you whether you will come back next week.

   What you should see: An empty project ready for instructions and files. The name reflects the use case.

3. Steps 3

   Add project instructions

   Open the three-dot menu in the top corner of the project and choose Project settings. In the project instructions field, write what ChatGPT must always remember here: role («You are an internal report editor»), tone («Direct, no empty adjectives»), output format («Bullets with owner and date») and bans («Do not invent figures; if data is missing, mark [MISSING]»). Project instructions apply only inside this project and can override your global custom instructions while you work here. Paste a sample paragraph if you have brand voice.

   What you should see: Project settings open with a 5–15 line block: role, tone, format and at least one explicit ban.

4. Steps 4

   Upload files or reference sources

   Use Add files or the project sources area to upload PDFs, spreadsheets, docs or images, or paste text the chats should consult. Do not upload the whole drive: 2–5 clear files are enough to start (total limits depend on your plan — Free allows fewer files per project than Plus or team plans). If a file goes stale, remove it. On plans that support it, you can link external sources your team updates; otherwise keep an exported copy current.

   What you should see: A file list with recognizable names. No duplicates or mixed stale docs.

5. Steps 5

   Open the first chat inside the project

   From the project, start a new chat — not from ChatGPT’s main screen. Write a concrete task that uses the context, with the same briefing rigor as chatgpt-primeros-pasos: «Using the project template, draft the March report. Use only data in the project files.» The chat inherits instructions and sources without you pasting them again. You can move older chats into the project if you already had a related thread.

   What you should see: A thread inside the project with your prompt. The reply follows the tone or template from instructions or files.

6. Steps 6

   Verify it uses the context (and fix if not)

   Review the reply: does it follow the format in the instructions? Does it use glossary terms? Does it mark [MISSING] instead of inventing? If something fails, edit project instructions in Project settings — do not repeat the same blind prompt. A «Banned: paragraphs longer than 3 sentences» in instructions beats three correction messages in the chat. If the problem is a stale file, replace it before continuing.

   What you should see: A reply aligned with instructions and files, or adjusted project instructions after spotting a failure.

7. Steps 7

   When to stay in the Project and when to return to a plain chat

   Stay in the Project for repeated work or a long thread on the same case: same client, same document type, same rules and sources that grow over time. Return to a plain chat for one-off questions, tone experiments or topics without shared files. If a second client needs another glossary and tone, create another project — do not mix contexts. If your team lives on claude.ai, the equivalent pattern is in claude-projects-primer-flujo; Notion AI can hold project notes, but chat with fixed instructions and files happens here in ChatGPT.

   What you should see: A documented decision: this project = Client X; other tasks = normal chat, another project or a GPT per chatgpt-gpts-cuando.