ChatGPT Space: helping the next shift pick up the incident

Imagine taking over an incident and finding the words “rollback complete.” Good. But are orders moving again? The answer may be somewhere in the chat, between an abandoned theory and a request nobody has answered. A useful handoff saves the incoming engineer that search without tidying away the uncertainty.

ChatGPT Space can hold that shared account: editable pages and files that people revise alongside the assistant. Instead of generating another summary for somebody to copy, a responder can add a finding to the document the next shift will read. OpenAI calls the product ChatGPT Space; individual spaces group related pages. The official overview explains the arrangement.

Picture a queue-backed order service after a deployment. Processing delays rise, the team rolls back, and the queue keeps growing. One engineer checks worker capacity; another investigates a dependency. This incident is invented, and the workflow below is proposed rather than tested. Its central problem is quite ordinary: an action is finished, but the question it was meant to answer remains open.

Leave the next engineer a place to start

The receiving engineer opens “Order processing delay: investigation and handoff.” At the top, a short account says what is known. Below it sit the alert export, deployment timeline and runbook reference, each with its time window. The timeline establishes when the rollback happened. The alerts show orders still waiting. Together they explain why someone is still investigating, not why the queue is growing.

In an account with Space enabled, New page opens a document that can begin with notes, files or an existing conversation. Related pages can share a space; one incident need not acquire a whole filing system. The getting-started guide describes the controls. Connected sources depend on account and workspace access, so approved, sanitized files are enough to try this without connecting the incident system.

A useful drafting request is specific about the unfinished parts: “Separate observed symptoms, completed actions, working hypotheses and unanswered questions. Keep source references and their time windows. Do not invent missing investigation results.” The next engineer needs the difference between what happened and what somebody suspects.

“Rollback resolved the delay” would sound reassuring. It would also erase the reason for the handoff. Keep the less comfortable, better-supported account. Later, the post-incident review needs that same distinction between a decision and what the evidence then available allowed people to conclude.

A payment check narrows the question

At 09:30, the dependency investigator returns: payment response times were normal in the sampled 09:10–09:25 interval. Now there is something useful to change. The payment-delay explanation has less support for those minutes. Orders are still waiting, worker capacity is still unexplained, and other dependencies have not somehow checked themselves.

The engineer can add the observation or ask ChatGPT to revise the affected paragraph. Selected-text edits and comments keep the review small enough to follow. Pages also offer child pages and reusable prompt blocks, but this particular finding needs a sentence, not another branch of the investigation. The working with pages guide describes these editing options.

When the next shift arrives, “payment response times normal, 09:10–09:25” gives it a useful starting point. “Dependencies ruled out” gives it an unjustified stopping point. The difference is only a few words, yet it decides whether the next engineer keeps looking.

One finding, one scoped change to the handoff

Read diagram description

A responder reports normal payment response times from 09:10 to 09:25. ChatGPT proposes a revision, and an engineer checks the source before accepting or correcting it. The handoff keeps the completed rollback, that limited payment check and the open worker-capacity investigation together. A separate, current service check must establish recovery.

Worker capacity passes to that engineer with a next check attached. If the investigation exposes a runbook gap, keep the proposed correction beside a link to the maintained procedure. It becomes the procedure in force only when the owner accepts and incorporates it. The SRE runbook guide describes what responders need from that authoritative version.

There is one more way to lose the thread: update Space while another responder updates the incident platform. Decide which owns current status and let the other support it. The incoming engineer already has a service to understand; two competing accounts add another problem.

Make sure the evidence opens too

First, can that engineer open the page? Available Space controls offer view, comment or edit access, with access inherited from a shared space or parent page. Sharing the whole space can expose more than the incident page. Removing a direct invitation does not remove access inherited elsewhere. The collaboration guide explains these distinctions and the account-dependent controls.

Then try the evidence link. Linked files keep their own permissions, so a readable handoff can point to an unreadable export. Copy private material into the page and its readers can see the copy; uploads follow page permissions. Private chats and saved memory are not shared merely by sharing the page, but details included in it are. The Space administration guide documents those boundaries.

Use synthetic or approved sanitized exports and ask the actual receiving engineer to open both the handoff and a source. A link that works only for its author is a small disappointment at a desk and a bigger one during an incident.

The page cannot watch the queue

Page revisions can come through the adjacent conversation, an inline request or a comment mention, depending on access. A dot, OpenAI’s personal AI assistant for work across connected apps, can also edit when it has access. Check whom a comment addresses: asking a person and asking an agent are different actions. The agents guide describes the routes and recommends reviewing the changes.

None of those editing controls watches the order queue. The finding entered at 09:30 describes measurements ending at 09:25. Keep both times. Those five minutes belong to the next check, not to the sentence already on the page.

OpenAI states that Keep Updated is unavailable at launch. Recurring work needs separate setup in chat, a saved schedule and an actual run to inspect. A page saying “refresh every five minutes” has not scheduled anything by itself. The same agents guide makes that limitation explicit.

The handoff can therefore say where the investigation stood at its latest reviewed observation. For recovery, the responder still looks at the service now: the queue and completed work, not merely the newly edited paragraph. The document saves reconstruction time so that check can happen sooner.

Choose the document the next shift will actually use

For this job, a Page holds the account and a space groups related pages and collaborators. A Project supplies shared chats, files, instructions and sources for related conversations; Sites builds hosted websites and applications. The official Projects and Sites guides describe those roles. A team may use several, but handing over a document does not require building an application.

Try one sanitized historical incident with an engineer who did not write the handoff. Notice where they pause, which source will not open, and which sentence brings them back with a question. Include the time spent reviewing generated edits. Those interruptions and that review work are part of the comparison, not an inconvenience to leave out.

The next shift should leave with three clear facts: the rollback happened, payment response times were checked for a stated interval, and worker capacity still needs attention. Use Space if it makes that account easier to maintain and inherit. If the incident platform already does the job, let the engineer read it there rather than maintain a second authority.