Skip to main content
The simplest way to avoid unnecessary credit use is to start a new conversation when the goal changes or you return after a long break.
When should you leave a long conversation behind? When you begin a different task, or return after more than an hour of inactivity—even if you are continuing the same work. The usual cache window is about one hour, so reopening an older conversation can mean paying for the model to process its full history again. Start a new conversation, or switch to a lower-cost model, use /compact, then return to the capable model if you need to continue the same task.

​
What should I do?

​
While actively working

​
Before taking a long break

Prepare the handoff while the cache is still warm. You can view or edit the saved project context in .workshop/context.md. If both apply, run /context first, then /compact.

​
When you come back

​
Three things that affect credit use

​
Why long conversations cost more

Each new turn includes the context needed to continue the conversation. As messages, file reads, and tool results accumulate, there is more history to process on every turn. A warm prompt cache makes repeated history cheaper, but it does not make the conversation smaller. Output, new input, agent turns, and delegated work still consume credits.
Why a fresh conversation saves credits. It leaves the old chat history behind while keeping the project’s files, sources, and saved project context. The agent starts with the durable information it needs, without carrying every earlier message and tool result.
Use one conversation for one clear purpose. When the purpose changes, start a new one.

​
Use /context to save knowledge for future conversations

/context saves important project knowledge from the current conversation in .workshop/context.md so future conversations can use it.
Run it after a milestone or lasting decision. Keep the saved context concise, and never include secrets. You can view or edit it in .workshop/context.md. See Context and Memory.

​
Use /compact to shorten history for the same task

/compact replaces a long working history with a shorter summary. Use it when you want to continue the same task but no longer need every intermediate step.
Compaction requires a summarization turn and usually starts a new cache. The saving comes from carrying the shorter summary on later turns. Workshop also compacts automatically near the context limit. See Working with the Agent.

​
Why caching and model switching matter

A prompt cache lets a model reuse conversation history at a discounted rate. The usual TTL is one hour. Think of this as an idle timer: using the cache can refresh it, but the exact behavior varies by model and provider.
Why a cold premium turn can be surprisingly expensive. Think of two prices multiplying. A premium model can cost around 10× more than a lower-cost model. If the cache is cold, rereading the same long conversation can cost around 10× more than reading it from a warm cache. Together, the conversation-history part of the first cold turn on a premium model can cost around 100× more than a warm turn on a lower-cost model.The exact difference varies, and the 100× comparison does not apply to the entire turn. The practical rule is simple: do not make a premium model process a long, cold history. Use a lower-cost model for /context or /compact, then use the capable model in the fresh or shortened conversation.

​
Choose the model for the task

The cheapest turn is not always the cheapest completed task. Repeated corrections can cost more than using a stronger model once. Auto is the simplest default when you do not want to pick a model yourself.