Chapter 04
Prompts
The updated instructions help Twig's support agent recognize the broken-amenity exception. Now Twig's engineers need to save those instructions, get them into the application, and verify which version produced each answer.
Save the corrected instructions
Twig's engineers use PostHog's prompt management to store instructions the application can fetch. They name the support prompt support-refunds.
The original cancellation instructions are v1. Twig's engineers publish the corrected instructions as v2. PostHog numbers each prompt version automatically and preserves its instructions and configuration, so the engineers can compare changes or return to earlier content.
Verify the version reaches the model
Twig's application is pinned to v1. Publishing v2 doesn't update that setting, so the support agent continues refusing the heating refund.
Send the complaint, then switch the application to fetch v2 and send it again. Inspect the version and instructions recorded with each answer:
Twig's engineers attach $ai_prompt_name and $ai_prompt_version to each generation. They also capture the instructions sent to the model. Together, those records connect an answer to the prompt that produced it.
PostHog's Usage tab shows which versions the application fetched by name. A fetch alone doesn't prove those instructions reached the model: if the generation still contains the old instructions, the engineers investigate the application's prompt fetch and cache.
Reuse a prompt with different inputs
The cancellation cases use the same instructions but different notice periods. Twig's engineers add {{days_before_arrival}}, a template variable: a placeholder filled with a value for each request.
Change the notice period to see what the application sends to the model:
The engineers also save their refund policy as refund-policy and include it through a prompt reference, which inserts another managed prompt's content. They keep that policy version fixed while comparing instructions, so a policy change can't explain a different answer.
They record the prompt and application-code versions with each offline experiment so another engineer can reproduce the comparison.
Inspect prompts in your project
Copy this prompt into PostHog AI: