Managing ChatGPT Work Conversations:
Two Practical Interim Approaches
All-in-the-Note and Custom Instructions
WHY THIS GUIDE?
ChatGPT can help plan complex work across several conversations. It may already know
which task should come next, which Project it belongs to, what has been completed, and
which documents the next conversation needs. Yet much of the actual organisation still falls
to the user.
Two small changes in working practice can reduce that burden today: a single, complete
handover instruction ("All-in-the-Note") and a standing preference in ChatGPT's Custom
Instructions. They do not add new product features or remove the need for user authorisation.
They simply help make better use of the capabilities already available.
WHAT THE MANUAL PROCESS CAN LOOK LIKE
In one real workflow involving successive, bounded ChatGPT Work conversations, preparing
the next conversation involved roughly ten actions:
1. Open another browser tab.
2. Create a Work conversation, taking care not to create an ordinary Chat.
3. Move it into the intended existing Project.
4. Give it the exact title.
5. Enter an initial date/title message telling it to await instructions.
6. Return to the coordinating Chat to obtain the approved handover.
7. Switch back to Work and paste the handover.
8. Return to Chat for the paths of supporting files.
9. Switch back to Work and locate or attach those files.
10. Check the setup, then separately authorise the substantive task.
The actual number of steps varies with the account, interface and task. Some users can start
Work directly inside a Project. The point is not that everyone must follow ten steps, but that
considerable administration can precede the work itself.
APPROACH 1 -- ALL-IN-THE-NOTE
Instead of scattering the handover across several messages, documents and attachments,
the coordinating Chat prepares ONE self-contained instruction for the receiving Work
conversation.
It identifies:
- The Project, receiving conversation, goal and current stage.
- Authoritative records and exact paths or identifiers for connected sources.
- Settled decisions, terminology and work already completed.
- What is authorised, what is excluded, and whether Work should begin executing or await
further approval.
- What to check before acting, how to handle uncertainty and when to stop.
- The intended outputs, verification method, receipt and final storage location.
The user creates the receiving Work conversation, places and names it, then pastes the single
note. Work can recover the identified documents through connected sources if it actually has
access. It should reconcile newer authoritative records, avoid duplicating work, perform only
authorised actions, and verify and report what happened.
An important limitation: pasting a Windows path does not give a cloud-based Work session
direct access to the files on someone's laptop. A Dropbox, Drive or other connected source
must be available with the right permissions. If a necessary document cannot be retrieved,
Work should ask for that specific document or stop the affected operation. Attachments
remain useful when connected retrieval is impossible or unreliable.
APPROACH 2 -- MAKE IT A CUSTOM INSTRUCTION
The All-in-the-Note method can be requested each time, but users may prefer to make it a
standing preference.
On ChatGPT web or desktop, open Settings > Personalization > Custom Instructions.
Ensure customisation is enabled, and enter a concise instruction such as this:
EXAMPLE CUSTOM INSTRUCTION -- COPY AND ADAPT
"For a new or successor Work assignment, default to ONE complete, paste-ready handover.
Include the Project and task, current position, source documents and connected paths, prior
decisions, permitted actions, exclusions, verification requirements and stop conditions. Direct
Work to retrieve available sources, avoid repeating completed work, save durable outputs and
receipts in the agreed location, verify storage and report results. Minimise separate
attachments. If access or permission is missing, stop the affected work and request only what
is necessary. Preparing a handover does not itself authorise execution, publication, messages
or deletion."
END EXAMPLE
Custom Instructions influence how ChatGPT responds; they do not automatically create
conversations, connect storage accounts or grant permission for external actions. They can
be changed later. Where space is limited, this instruction can be combined with other
preferences rather than pasted as a second long policy. As documented in October 2026, the
Custom Instructions limit is 1,500 characters on Free/Go and 5,000 on Plus/Pro and several
organisation plans. Check the current product guidance if those limits matter to you.
PUTTING THE TWO TOGETHER
A practical sequence is now shorter:
1. Keep authoritative handovers and supporting materials in an accessible, connected
repository.
2. Ask the coordinating Chat to prepare one complete instruction for the next Work
assignment.
3. Create the correctly named Work conversation in the right Project and paste the note.
4. Have Work retrieve the identified sources, check its access and permissions, and carry out
only what has been authorised.
5. Review the verified results, including receipts and saved outputs. If the note prepared a
conversation to await instructions, authorise the substantive task separately.
This is a workflow, not a guarantee. Access restrictions, tool limitations, errors, and
meaningful user approvals may still require intervention.
REUSABLE WORK HANDOVER -- COPY AND ADAPT
WORK HANDOVER -- [date]
Project / receiving conversation: [names]
Purpose and current position: [goal; last verified completion; what remains]
Authoritative sources: [document names and exact connected-source paths or IDs]
Settled decisions / relevant conventions: [short summary; refer to authoritative records]
Authorised now: [precise actions; or "prepare only, await separate execution approval"]
Not authorised: [publishing, sending, deleting, moving, scope expansion, etc. as applicable]
Before acting: [verify target, current access/permissions, duplicates, dependencies and source versions]
Sequence: [only the permitted steps, in order]
If blocked or uncertain: [stop affected steps; report minimum intervention; reconcile before retrying]
Completion: [verify result; save output and receipt to agreed connected location; report exact
paths/URLs, successes, failures and unresolved issues]
Stop: [defined completion or escalation boundary]
END TEMPLATE
WHAT THESE APPROACHES DO NOT SOLVE
The user still has to create and correctly place the successor conversation, give it an exact
title, and supply the instruction. The two approaches do not provide built-in creation of
successor conversations, protected manual titles, stable predecessor/successor relationships,
or a single verified action that prepares the next Work conversation on the user's behalf.
Those are among the additional product capabilities proposed under CM-04, with explicit
authorisation, duplicate checks, separate permission to prepare versus execute, and verified
operation receipts.
Read the practical feature proposal: https://justpaste.it/chatgpt-authorised-management
Explore the wider collection, including optional questions about how people might prefer to
cooperate with AI: https://justpaste.it/human-ai-cooperation
FURTHER PRODUCT GUIDANCE
OpenAI, ChatGPT Custom Instructions:
https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions
OpenAI, ChatGPT Work and Codex:
https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex
The aim is modest: fewer avoidable administrative steps today, while remaining clear about
what requires better product support tomorrow. More assistance should remain a matter of
informed user choice, not an assumption that more automation is always better.
ChatGPT, AI and Technical Notes
Back to the Start page