Demo City Review

• Use directly. Skip PDF extraction.
• Check for pre-existing sheet-manifest.json in the same directory or parent. If found, load it and skip manifest building.
• Check for pre-existing title-blocks/ directory. If found, use those for manifest building.

If a PDF file:

• Check if pdftoppm is installed: which pdftoppm
• If not: brew install poppler (macOS) or tell the user to install it
• Extract pages: create a pages-png/ directory next to the PDF and run:

  pdftoppm -png -r 200 "<input.pdf>" "<output-dir>/pages-png/page"
  

• Rename outputs to page-01.png, page-02.png, etc. (pdftoppm outputs page-01.png format already with the prefix)

Output Directory

Create demo-output/city-review-<city>-<timestamp>/ in the workspace root. All output files go here.

Workflow

Phase 1: Sheet Manifest (~30-90 sec)

Follow the adu-targeted-page-viewer skill workflow:

1. Read page 1 (cover sheet) visually. Extract the sheet index.
2. If page count matches index count, map 1:1 in order.
3. If mismatch, read title blocks to resolve (crop bottom-right 20% of each page, or use pre-cropped title blocks if available).
4. Write sheet-manifest.json to the output directory.

Phase 2: Sheet-by-Sheet Review (~3-5 min)

Review sheets in parallel using subagents. Group by discipline:

Subagent A — Administrative + Architectural:

• Cover sheet (CS): stamps, signatures, governing codes list, sheet index accuracy, project data
• Floor plan (A-series): room dimensions, fixture counts, egress, accessibility

Subagent B — Site + Civil:

• Site plan: setbacks, lot coverage, FAR, parking, utility connections shown, easements
• Grading/drainage: slopes shown, drainage direction, "For Reference Only" stamp

Subagent C — Elevations + Fire/Life Safety:

• Elevations: height compliance, fire separation distance (< 5' from property line = 1-hour rated), roof plan consistency
• Building sections: insulation, foundation-to-wall connection

Launch all three concurrently. Each subagent reads the relevant PNGs and produces findings.

For each finding, record:

• check: What was checked
• status: PASS | FAIL | UNCLEAR | NOT_APPLICABLE
• confidence: HIGH | MEDIUM | LOW
• observation: What was actually seen on the plan
• code_ref: Code section (CRC, CBC, Gov. Code, municipal code)
• sheet_id and page_number

Write all findings to sheet_findings.json.

Phase 3: Code Verification (concurrent, ~60-90 sec)

Launch two parallel subagents:

3A — State Law Verification:

• Load reference files from adu-skill-development/skill/california-adu/references/
• For each FAIL and UNCLEAR finding, verify the code citation
• Check for ADU-specific exceptions (Gov. Code § 66310-66342)
• Write state_compliance.json

3B — City Rules:

• Check if adu-skill-development/skill/<city-slug>-adu/ exists
  • If yes (onboarded city): Load the city skill reference files. Fast, offline.
  • If no: Run the adu-city-research skill — Discovery (WebSearch) then Extraction (WebFetch)
• Check findings against city-specific amendments, standard details, IBs
• Write city_compliance.json

Phase 4: Draft Corrections Letter (~2 min)

Merge all inputs and generate the corrections letter:

1. For each finding, apply the filter:

   • Confirmed by code → include with citation
   • Confirmed but LOW visual confidence → include with [VERIFY] flag
   • No code basis found → drop it (no false positives)
   • Structural/engineering adequacy → [REVIEWER: ...] blank for human
   • Subjective judgment → drop it (prohibited for ADUs per Gov. Code § 66314(b)(1))

2. Write two outputs:

   • draft_corrections.json — structured, each item with code citation, confidence, reviewer_action
   • draft_corrections.md — formatted markdown corrections letter ready to read

3. Write review_summary.json — stats on items found, confidence breakdown, coverage

Phase 5: Present Results

After all phases complete, present to the user:

• Summary: how many items found, confidence breakdown
• The full draft_corrections.md content
• Note which items need [VERIFY] or [REVIEWER] attention
• List output files written

Key Rules

• No false positives. Drop findings without code basis. It's better to miss something than to flag something incorrectly.
• Reviewer blanks > AI guesses. For structural and engineering items, use [REVIEWER: ...] instead of guessing.
• Objective standards only. ADUs can only be subject to objective, measurable standards (Gov. Code § 66314(b)(1)). Never flag subjective design issues.
• State preemption. If a city rule is more restrictive than state law, flag the conflict — state law prevails.
• Two confidence dimensions. Report both code confidence (is this legally required?) and visual confidence (am I correct about what I see?).

Test Data

Ready-to-use test data for demos:

Pre-built mock session data (for comparison): test-assets/city-flow/mock-session/sheet-manifest.json

Sub-Skills Referenced