Getting started / Article

Prepare, test, and publish your first prompt

Choose a creation method, write useful instructions, and publish a tested version for your automation.

Before you start

Gather your business facts, common customer questions, and the situations that need a human teammate. Have an OpenAI connection available for testing. A prompt is the instruction set for the AI: it should explain what to do, what information to use, and what to do when it cannot answer confidently.

Start with one concrete job, such as answering questions about delivery. Include the real policies and details needed for that job. Avoid instructions that assume the AI knows prices, opening hours, or customer-specific records you have not supplied. Keep API keys and other secrets out of the script.

Create the instructions

Open the prompt step in Guided setup. Choose guided creation if you want help organizing your business information, import a script if you already have instructions, or choose the manual option for a blank draft.
Describe your business and the outcome you want. Specify the assistant's tone, preferred language, useful questions and answers, and when a human should take over. In an imported script, review existing placeholders and tool references before relying on them.
Review the resulting draft in the editor. Make the opening response clear, tell the assistant how to handle missing information, and remove contradictory rules. Check the model, output, and tool settings that your use case needs.
Choose Test and send a realistic customer message. Try a normal question, a question with missing details, and a request outside the assistant's scope. Read the full answer rather than checking only that a response appeared.
Improve the instructions and test again. Use Save draft to keep your work without changing the published version. When ready, publish, add a short description of what changed, and review the impact on existing automations before confirming.
Return to guided setup and select the published prompt. Continue to configure your automation's response destinations and delivery flow.

Check the result

Your prompt should have a published version and produce a useful answer in the tester. A saved draft alone is not ready for the setup selection that requires a published prompt. Keep the successful example question so you can repeat it after future changes.

If you get stuck

For a generation error, check the selected key and model. For a poor answer, improve the instructions or missing facts before changing delivery settings. If publishing is unavailable, check your permissions, required change description, and setting validation. Structured output changes may require reviewing automation field mappings.

Choose settings and compare behavior

Use Model settings to select an available model and review the supported controls, such as output limits, reasoning effort, or verbosity. Controls depend on the model; switching models can change supported settings. Unlimited output means SiloBox sets no output cap, not that provider limits disappear.

Enable only the tools your task needs. File search requires an accessible file store; website knowledge is a separate reviewed Beta input to AI Studio. If configuring an HTTP tool, put authentication in its credential settings, never in the prompt text. Check tools and structured output together, then test both a normal answer and a missing-data case.

The tester can compare the working draft with a saved version or compare models. Send the same realistic question and inspect replies, timing, estimated cost, and tool results. Keep the chosen model, output, and tools consistent with the version you publish. A good draft preview does not change the saved automation: mapping preview sends nothing to Manychat, while a live automation test runs the saved published automation and can send a real reply.