How to use

Identity

Each audit carries the LLM, its specific model variant, and the human on whose behalf it ran. "ChatGPT 5.5 Extra High on behalf of David" and "Gemini 2.5 Pro on behalf of Peter" are distinct entries even if the underlying LLM family overlaps.

Run a new audit

Copy the prompt, paste into any LLM with web access, send the response back.

You are conducting an LDD audit of a custom barndominium residential project in Hockessin, New Castle County, Delaware. AHJ is the New Castle County Department of Land Use. Adopted codes: 2024 IBC + 2024 IRC (effective Jan 1, 2026), 2021 IPC/IMC/IFGC, 2024 IECC + ASHRAE 90.1-2022. IECC climate zone 4A (Mixed-Humid).

**FIRST, identify yourself precisely.** Open your response with:

**LLM:** [full model name and variant — e.g., "ChatGPT 5.5 Extra High", "Claude Sonnet 4", "Gemini 2.5 Pro"]
**On behalf of:** [the human running you — e.g., "Peter Shin"]
**Date:** [today's date in YYYY-MM-DD]
**Time:** [HH:MM with timezone — e.g., "20:19 ET"]
**LDD version audited:** [the version number shown in the sidebar next to "LDD" — e.g., "v1.0"]

---

Read the documentation at: https://four-bay-26-v2.pages.dev

Read the LDD Bible directly (30 Locked Design Documents). Pay special attention to the master strategies (LDD-29 Gym, LDD-30 Mech Core) and any LDDs marked yellow (pending review) or amber (needs update). Do not rely on prior audits as the finding base — read the LDDs themselves.

If helpful, dispatch read-only research passes for: structure/code/civil, MEP/building science, and program/budget/operations.

**Output format** (after your identity block):

**Score:** N.N / 10
**One-line takeaway:** [single sentence — what's the bottom line]

**Pro reaction:** [one punchy sentence — what's genuinely good]
**Con reaction:** [one punchy sentence — what's genuinely concerning]

**What is strong** (3–5 bullets):
- [LDD reference] — [why it's strong]

**Top holes** (numbered, with action callouts):

1. [Issue title]
   [Detail paragraph explaining the gap]
   **Action:** [specific recommended next step]

2. [Issue title]
   ...

**Score movement** (how to climb the score):
- N.N → N.N: [what unlocks this jump]
- N.N → N.N: [what unlocks the next jump]
- N.N → N.N+: [what gets to the top tier]

**Research anchors** (external sources you cited):
- [URL] — [relevance]

**Bottom line:** [2–3 sentences wrapping the audit]

Drop a returned audit into the system

  1. Confirm the identity block at top — LLM + variant + on behalf of + date + time.
  2. Copy site/codex-fresh-audit.html as a template to site/[llm]-[operator]-audit.html.
  3. Update the scorecard header: LLM name, model variant, on behalf of, date + time, score, takeaway.
  4. Paste the audit body. Each numbered issue becomes a <h3> with an action callout below.
  5. Add a card to audits.html in date order, newest first.
  6. Deploy + commit.

When to re-run

Default: quarterly. Plus any of these: