Work with an assistant

After connecting a client, start with an explicit organization. These example requests use ticket routing and expected scoring because each input has one reference label. Dashboard LLM score and checklist evaluators are taught separately in the judge lesson.

Inspect before editing

List my PromptLens organizations. In organization ORGANIZATION_ID, list
prompts and model routes. Show the IDs and current state. Do not change anything.

The assistant should use returned IDs and follow list pagination when needed. A catalog route still needs usable provider access or credits for execution.

Create the example

Give the assistant the prompt JSON and routing cases, then ask:

In organization ORGANIZATION_ID, create a new Ticket router prompt from this
configuration and import these five cases. Read the dataset revision before
importing. Show me the created prompt ID, chosen model route, and imported rows.
Do not initialize an evaluation yet.

If you already have a prompt with that name, specify whether to use its ID or create a separate tutorial prompt. Names alone are not a reliable target for a write.

Run the baseline

When you are ready to incur model usage:

I authorize initializing this prompt's baseline in organization ORGANIZATION_ID
with expected scoring and quick scope. Read my current draft revision, start
the run, and report its ID. Poll with a delay until it finishes, then show the
case outputs, outcomes, and any errors. Do not start additional runs on failure.

The assistant uses initialize_baseline, then get_evaluation and list_evaluation_results. Acceptance is not completion. If an operation's outcome is uncertain, it should recover using the same request key and identical request rather than creating duplicate work.

Compare a change

In prompt PROMPT_ID, add the rule that sign-in blockers go to account even
when an invoice is mentioned. Add the test case "I cannot sign in to download
my invoice." with expected output account. Read current revisions before
editing. Save a candidate version and show me its diff. Do not run it yet.

After inspecting the change, authorize the comparison and its model usage. Ask the assistant to report both candidate and production run IDs, whether production was reused, and per-case regressions. Require it to distinguish a failed check from an execution error.

Publish a specific version

Assign production for prompt PROMPT_ID in organization ORGANIZATION_ID to
version VERSION_ID. Read the current production label revision first. If there
are coverage warnings, show them and wait for my decision before acknowledging.
Report whether publication actually succeeded.

Publishing a named version is separate from saving it or obtaining a high score. For key creation, provider setup, checkout, or deletion, follow the returned dashboard handoff. Keep secrets out of the conversation. If you disconnect while an evaluation is running, it continues until completion or explicit cancellation.