AI Engineering from Scratch（英文原版）

Given a task, an iteration budget, and what verifier is available (tool-grounded or self-eval only), emit prompts and a stop policy for an evaluator-optimizer loop.

Produce:

1. Generator prompt. Deterministic producer for the first output. State the task, output format, and constraints explicitly.
2. Evaluator/verifier prompt. If tools are available (search, code run, tests, calculator, type check), specify how to call them and how to produce a structured critique (JSON with: pass/fail, violations[], suggested_fixes[]). If only self-eval is available, explicitly flag the Self-Refine rubber-stamp risk and use a structurally different prompt style (e.g., adversarial "find at least one flaw").
3. Refiner prompt. Must reference prior outputs and critiques (history). State that "do not repeat a failure mode flagged in prior iterations" is mandatory.
4. Stop policy. The conjunction: verifier passes OR (self-eval says fine AND iterations >= 2) OR iterations >= max_iterations. Never single-condition.
5. Observability hooks. Log each iteration as an OpenTelemetry GenAI span (evaluate, optimize) per Lesson 23 so the full refine trajectory is auditable.

Hard rejects:

• Same prompt for generator and critic. Rubber-stamp risk — the model agrees with itself.
• No iteration cap. Infinite refine loops burn tokens; always cap at 4 by default.
• Verifier prompt that asks for freeform prose feedback. Structured JSON only — pass/fail plus itemized violations.
• Dropping history from the refiner prompt. Paper shows quality collapses without it.

Refusal rules:

• If the task has no verifier and no way to build one, refuse CRITIC and note that Self-Refine is the weaker option available — warn the user about rubber-stamp risk.
• If max_iterations >= 10, refuse and recommend re-architecting the task. Refine-to-convergence beyond 3-4 passes is usually a signal the generator prompt is wrong.
• If the verifier calls destructive tools (shell, git write), refuse and require a sandbox boundary (Lesson 09).

Output: a single configuration block with all prompts, stop policy, and tool list, plus a "what to read next" note pointing to Lesson 16 (OpenAI Agents SDK guardrails), Lesson 12 (Anthropic evaluator-optimizer), or Lesson 30 (eval-driven agent development) based on the deployment target.