System Prompt Workspace¶
Route: /#/basic/system
Use this workspace for long-lived constraints such as role design, behavior boundaries, and output policy.
If your main problem is how to phrase one concrete task, User Prompt Workspace is usually a better fit.
First-time rule of thumb¶
If both are true, this is usually the right page:
- you are editing long-lived role / rule / boundary content
- the right side needs one test message before it can run
Typical use cases¶
- define a stable model role
- enforce a reusable output structure
- define what the assistant must not do
- test whether the model keeps rules across different questions
If you only want the fastest start¶
- enter one system prompt on the left
- run one left-side optimization
- enter one test message on the right
- run testing, then Result Evaluation
What the left side edits¶
The left side edits the system prompt itself.
You can think of the page like this:
- upper-left: original system prompt
- lower-left: current workspace draft and saved versions
What the right side tests¶
The right side tests:
- one system prompt version
- one fixed test message
- the real output
That is why the right-side test message is required in this workspace.
Analysis vs evaluation in this page¶
- left-side Analysis: inspects the system prompt itself, not the test message
- right-side Result Evaluation: judges whether one real output reached the goal
- right-side Compare Evaluation: compares multiple real outputs
Recommended workflow¶
- enter the original system prompt
- optimize it once on the left
- use left-side analysis if you want prompt-only feedback first
- enter one fixed test message on the right
- compare
original / workspace / vN - start with Result Evaluation
- then run Compare Evaluation if you have multiple columns
- apply valuable suggestions back to the left workspace
Common confusions¶
- Left-side analysis does not read the test message. It analyzes the system prompt itself.
- The right-side test message is required because a system prompt usually cannot reveal its behavior on its own.
A minimal example¶
System prompt:
You are a customer support assistant.
Right-side test message:
My order has not shipped for three days. Can I request a refund now?
With that setup, you can compare:
- whether the original version is too vague
- whether the workspace version follows boundaries more reliably
- whether different models misunderstand the same system prompt in different ways