Module 03 Activity

Scenario

The next step after your model call is code, not a person. Everything here is about making the boundary safe.

What you build

A parser with format instructions, a validating schema, a feedback-driven retry with a give-up path, and a schema-change plan.

Steps

1. Replace prose output with a parser, binding the format instructions with .partial() so they cannot be forgotten.
2. Run four candidate outputs through the parser, including JSON wrapped in prose and truncated JSON. Note which succeed.
3. Add a validating schema constraining types, allowed values and at least one format pattern.
4. Build a retry that feeds the specific error back, caps attempts, and has an explicit give-up path.
5. Count parse failures separately from validation failures and alert on each rate.
6. Write the schema-change plan: which changes are backward compatible and what your rolling-deploy order is.

Evidence to hand in

• The parser with bound format instructions.
• Four candidates with their outcomes, including the two surprises.
• The schema with types, allowed values and a pattern.
• The retry with its feedback, cap and give-up path.
• The schema-change plan with deploy ordering.

Review checklist

• Format instructions are bound once, not passed per call.
• The schema rejects a plausible label outside the allowed set.
• The retry feeds back the specific error, not 'try again'.
• There is an explicit give-up path.
• The deploy plan says readers before writers for a widened enum.

Mark complete