Save a preference when it is stable, useful and appropriate to retain. Keep a one-off project fact in the task where it belongs. Before enabling memory, decide what future conversations should be allowed to reuse.
Custom instructions, memory, project context, training preferences and deletion controls do different jobs. Similar labels across products do not establish identical behavior, access or retention.
Put the context in the right place
| Material | A suitable starting point | What to check |
|---|---|---|
| Stable response preference | Custom instructions | Is it a default or an absolute that will obstruct other tasks? |
| Reusable low-risk personal context | Optional memory, if you want personalization | What is retained, where it comes from and how to correct it |
| One project’s instructions and sources | An authorized project/workspace | Scope, members, source versions and sharing |
| A temporary task fact | The current task | Whether it needs to persist at all |
| Confidential or sensitive material | Only an explicitly authorized data path | Permission, necessity, access and retention before submission |
A project folder is not proof of a security boundary. A work email address is not proof that every data class is approved. Start with fictional material while learning the controls.
Set two defaults you can test
For example:
Use short paragraphs unless the task needs a table or another format.
Preserve supplied facts and uncertainty. Do not invent missing dates, locations
or commitments; ask when a missing detail prevents completing the task.
These are response instructions, not facts about your life. They avoid storing an employer name, client details or personal history merely to obtain clear prose. OpenAI’s Custom Instructions documentation explains its feature; use the current controls in your own account rather than relying on an old screenshot.
Do not write “never hedge” beside “be accurate about uncertainty.” Prefer “remove filler while preserving qualifications.” If a default damages a particular task, override it explicitly and inspect whether the result changed.
A missing-venue test
This is an authored teaching example, not a live test of saved ChatGPT settings.
Source notes: The workshop is on Friday. The venue is undecided.
Task: Write a short invitation using only these notes.
Checked output:
“Join us for the workshop on Friday. The venue has not been decided yet.”
This preserves the known day and unknown venue. “Join us at the studio at 10” invents two details. “We will send the address tomorrow” invents a commitment, even though it sounds helpful.
Now test a different format: ask for a one-line calendar description. The short-paragraph default should not prevent a one-line answer. Then start an unrelated fictional task and check that the workshop context has not leaked into it. Those are useful observations; one successful result does not prove every future answer will follow the defaults.
Inspect memory separately
OpenAI’s Memory documentation describes personalization and its available controls. Inspect what your account currently exposes, correct stale context and follow the current removal instructions for the sources involved. A summary of remembered context should not be treated as proof that you have inspected every retained chat, file or connected source.
Turning personalization off, deleting a conversation and switching off model-training use are separate actions. Read the Data Controls FAQ for training settings. Do not put sensitive data into memory simply because you found a training toggle.
Temporary Chat also has its own modes and retention rules. Check whether the mode you select uses existing personalization and what happens if you save it. Do not interpret “temporary” as no processing, no possible retention or permission to share confidential information.
Keep projects maintainable
Use project instructions for a coherent task whose sources you are allowed to share. Claude’s Projects documentation explains its product’s project context; verify visibility and membership in your actual workspace before adding files.
Record which facts are current, which source supersedes another and who maintains them. Remove or correct stale context when a role, project or policy changes. If an instruction is repeatedly ignored, reconsider the workflow rather than growing an untested wall of rules.
Use the context exercise to compare a plain prompt with one carrying your defaults. Keep what helps, and retain only context whose future reuse you actually want.



