Build And Verify

<!--
  GENERATED FILE — do not edit by hand.
  Source of truth: client/src/lib/content.ts (BUILD_VERIFY_MARKDOWN)
  Regenerate: pnpm generate:prompts
  Agent Handoff Framework by Roy McFarland (github.com/roymcfarland).
  License: MIT (same as the repository) — vendor freely.
  Guide: https://www.worksmithlabs.com/
-->
# Build & Verify — Two-LLM Workflow

A disciplined loop for shipping with AI agents: an Advisor scopes each
slice and drafts both prompts, one LLM writes the code, and a *different*
LLM verifies it in a clean context window before the next slice runs.

---

## Why two LLMs?

Code-capable models can produce persuasive summaries of work they authored.
Empirical LLM-as-judge work shows judges vary by prompt, model, and task (e.g.,
Zheng et al., https://arxiv.org/abs/2306.05685). A Verifier LLM in a fresh context,
reading only the PR description,
diff, and scope document, is a structured second reader that can surface stubs,
speculative APIs, out-of-scope edits, and silently skipped Acceptance Criteria
that the Builder may under-report — especially when verdicts must cite evidence.

Two-model separation does **not** guarantee independence; it raises the odds
that blind spots differ from the Builder's. Narrow inputs and explicit verdict
formats keep that signal usable.

---

## The loop