How to set it up
How prompts work in Oodle
Manage, update, edit, compare, release.
1 · Manage
Keep every prompt in Oodle. Your agent fetches it at runtime.
Your code asks for a prompt by name and label, fills in the variables, and runs. The prompt text lives in Oodle, not in your repo, so changing it never means a deploy.
- 1# Fetch by label
- 2prompt = oodle.get_prompt(
- 3 "voice-agent-appointment-booking",
- 4 label="production",
- 5)
- 6
- 7# Compile with variables
- 8compiled = prompt.compile(
- 9 caller_name="Priya",
- 10 transcript="Hi, can I move my Friday appointment to next week?",
- 11)
2 · Update
Update prompts directly from Oodle.
Every edit is a version. Labels like production and candidate point at one version each, so releasing a change is moving a label.
- v8candidatetoday
- v7production3 days ago
- v62 weeks ago
- v51 month ago
production label
v7 → v8
Your code fetches by label. Moving it is the release; there is nothing to deploy and it is one click to move back.
Promote v83 · Edit
Edit the prompt next to the conversation that broke.
Open a failing session with its inputs already filled in. Change the prompt, swap the model, try again. Nothing to deploy.
4 · Compare
Replay it against real conversations.
Keep the sessions that went wrong as test cases, real ones, not made up. Replay every candidate against them and score both versions with the evals you already trust.
5 · Release
Make the winner live. Keep watching.
Move the production label to the version that scored best. The evals keep scoring production, so if the problem comes back you see the drift before customers tell you.
“Your return has been filed.” — it wasn’t. initiateReturn returned 503.returns-failures · 24 conversations like this one
Step 1 · Failing conversations
Real conversations, kept as test cases.
Now change the prompt.