ChatGPT Projects: organize an ongoing work project (2026)

If you already built your first Project with chatgpt-projects-primer-flujo, the next problem is different: work runs for weeks, sub-tasks multiply, stale files pile up and threads nobody reopens. ChatGPT Projects groups chats, files and instructions for one body of work (official guide: https://help.openai.com/en/articles/10169521-projects-in-chatgpt). This flow does not repeat New project or the first chat — it assumes a live project and gives you rules to organize it without mixing contexts or losing decisions. About 25 minutes of setup plus a short weekly habit.

1. Steps 1

   Sketch phases and sub-tasks before opening more chats

   In a notes app (outside ChatGPT), list project phases — for example Discovery, Draft, Legal review, Close-out — and under each phase write 2–4 concrete sub-tasks. That list is your chat naming map: one thread per active sub-task, not one mega-thread for the whole quarter. If a sub-task already has a loose chat in general history, move it into the project when Move to project is available (drag onto the project or use the chat menu); if the chat is not eligible, start a new one inside the project and paste only the summary that matters. Write the project deliverable and target date in one line.

   What you should see: A phase diagram with sub-tasks and a written rule: «One chat = one active sub-task.»

2. Steps 2

   Structure project instructions in two layers

   Open Project settings (project ••• menu). In the project instructions field, keep two mental blocks in one place: (1) Stable — role, tone, deliverable format and bans («Do not invent figures; mark [MISSING]»), what does not change every week; (2) Operating — how to work here over weeks: «Name each chat Phase — sub-task», «After a meeting, output decisions as dated bullets», «Cite filenames when requesting drafts». Project instructions apply only inside this project and override your global custom instructions while you work here. Do not replay the whole chatgpt-projects-primer-flujo setup — only add the operating layer a long project needs.

   What you should see: Project settings with 10–20 lines: stable block on top and filing/naming rules below.

3. Steps 3

   Curate files: canonical, draft and out of project

   Review the project sources list. Mark as canonical only what ChatGPT should always consult: glossary, approved template, signed brief, reference dataset. Week-specific drafts can live in a chat or in a doc you update; when a draft becomes final, replace the old file or delete it from the source preview (deleting removes it from the project for everyone on a shared project). Avoid accidental duplicates: on upload, ChatGPT offers Upload anyway or Skip when a filename already exists. Use explicit filenames (client-template-v3.pdf). If material grows large, consider whether heavy document study fits notebooklm-primer-briefing first and keep in the Project only what you need to draft or decide.

   What you should see: 4–8 sources with clear names; no stale copies; your own note of what is canonical.

4. Steps 4

   One chat per sub-task; branch to explore

   Create or rename chats inside the project using your pattern — for example «Review — vendor contract». Work the sub-task only in that thread until you close or pause it. To try a different approach without losing the original chain, use Branch chat on an existing thread: the branch appears in the project next to the original. For one-off questions without shared files, use a plain chat; for reusable roles in short sessions, read chatgpt-gpts-cuando before duplicating logic in another container.

   What you should see: Several chats with readable names; at most one or two active branches per exploratory sub-task.

5. Steps 5

   Keep a living decisions and summary log

   Maintain a master doc — for example project-decisions.md — with sections: Decisions (date, who, what was chosen), Open items and Pointers to key chats. Each week, in the active sub-task chat, ask: «Summarize only new decisions since [date] as dated bullets; do not rewrite the full history.» Edit the output if it invents or softens choices, then update your master. When a reply is good enough to reuse as fixed reference, open the message menu and choose Save to project / Add to project sources (label may vary) so future project chats can draw on it. That beats scrolling a 200-message thread.

   What you should see: An updated decisions file and optionally one or two responses saved as project sources.

6. Steps 6

   Set project memory to the boundaries you need

   In Project settings, under Memory, choose Default memory or Project-only memory and save. With project-only memory, project chats can reference other conversations within the same project, but not conversations outside it or globally saved memories; project information also stays out of chats outside the container. With default memory, behavior depends on your plan and personal settings (Enterprise/Edu workspaces keep project chats more contained — see the same official guide). Shared projects always use project-only memory. For sensitive client work, project-only memory is often the clearer default; if you want the project to use personal tone memories, consider default and review chatgpt-memoria-personalizacion. Changes may take a few hours to apply.

   What you should see: Project settings with Memory set to Project-only or Default and the choice noted in your log.

7. Steps 7

   Clean up, archive or close when the phase shifts

   Every two or three weeks (or at phase close), inventory: delete stale sources, capture what still matters in your log and archive or delete chats that no longer add context — on shared projects, members can archive or delete chats they own per the official guide. Rename paused chats with a «[Done]» prefix so you do not reopen them by mistake. When the effort is finished and you will not return, choose Delete project from •••: that permanently removes the project’s chats and instructions and files stored only in the project (files saved separately in Library are not deleted). This cannot be undone. If you only pause for months, refresh operating instructions with «status as of [date]» and keep the decisions log current instead of deleting the container.

   What you should see: A shorter chat list, sources without old duplicates and a written choice: keep, pause or delete project.