ADU Targeted Page Viewer

Check first: PNGs may already be pre-extracted. Look for pages-png/page-01.png in the project files directory. If PNGs already exist, skip this step entirely — go straight to Step 2.

If PNGs don't exist, run scripts/extract-pages.sh to split the PDF into individual page PNGs:

scripts/extract-pages.sh <input.pdf> <output-dir>

This produces output-dir/pages-png/page-01.png, page-02.png, etc. at 200 DPI (full resolution, no resize). Takes ~5 seconds for a 26-page set. The script is idempotent — if PNGs already exist it exits immediately.

Step 2: Read the Cover Sheet and Find the Sheet Index

Read pages-png/page-01.png visually. The sheet index is typically in the top-right or right-side area of the cover sheet.

Extract the index as a list of entries:

CS    → Cover Sheet
A1    → Site Plan
A2    → Floor Plan
A3    → Elevations & Roof Plan
S1.0  → Structural Notes
S2.0  → Foundation Plan
...

This gives the list of expected sheets and their descriptions.

If the index is not on page 1, check page 2 — some sets have a separate title page before the cover sheet.

Step 3: Match Sheet IDs to Page Numbers

The sheet index order generally matches the PDF page order, but there can be mismatches — the index might list 12 sheets while the PDF has 15 pages (extra city forms, checklists, etc.).

To resolve the mismatch, read the title blocks:

1. Check if title block crops already exist at title-blocks/title-block-01.png. If they do, skip the cropping step. Otherwise, run scripts/crop-title-blocks.sh to crop the bottom-right corner of each page:

scripts/crop-title-blocks.sh <output-dir>/pages-png <output-dir>/title-blocks

2. Read each title block image to extract the sheet ID. Title blocks are small crops, so this is fast — the sheet ID is prominently displayed.

3. Match each page's sheet ID against the index entries.

Optimization: If the page count matches the index count exactly, skip the title block reading — the index order IS the page order. Only do title block reading when there's a mismatch.

Parallel processing: When reading title blocks, launch parallel subagents (3-5 at a time) to read batches of title block crops. Each crop is tiny, so reading is fast.

Step 4: Build the Sheet Manifest

Output a JSON manifest file:

{
  "source_pdf": "plans.pdf",
  "total_pages": 15,
  "indexed_sheets": 12,
  "sheets": [
    {
      "sheet_id": "CS",
      "page_number": 1,
      "file": "page-01.png",
      "description": "Cover Sheet"
    },
    {
      "sheet_id": "A1",
      "page_number": 5,
      "file": "page-05.png",
      "description": "(E) & (N) Site Plan"
    },
    {
      "sheet_id": "S2.0",
      "page_number": 11,
      "file": "page-11.png",
      "description": "Foundation Plan"
    }
  ],
  "unindexed_pages": [
    {
      "page_number": 2,
      "file": "page-02.png",
      "title_block_text": "City of Long Beach Forms"
    }
  ]
}

Save this as output-dir/sheet-manifest.json.

Step 5: Targeted Sheet Viewing (On Demand)

When a corrections letter references a specific sheet:

1. Look up the sheet ID in the manifest → get the page number and PNG path
2. Read that PNG visually
3. Analyze the specific area referenced (e.g., "Detail 2" → look for the detail numbered "2" on the sheet)
4. Report what is on the sheet and what needs to change

See references/plan-sheet-conventions.md for how to navigate construction plan sheets — where title blocks are, how detail callouts work, and the sheet numbering system.

Integration with Corrections Interpreter

The corrections interpreter extracts sheet references from the corrections letter during its Phase 1 (reading the letter). It passes those references to this skill:

Corrections interpreter: "I need sheets A1, A3, S1.0, and S2.0"
    ↓
This skill: "A1 = page-05.png, A3 = page-07.png, S1.0 = page-09.png, S2.0 = page-11.png"
    ↓
Corrections interpreter reads those PNGs and produces targeted findings:
    "On Sheet A3 (page 7), Detail 2 in the lower-right shows a standard soffit.
     The patio is within 5' of the property line — replace with 1-hour fire-rated
     assembly detail showing 5/8" Type X gypsum on the underside."

Important Notes

• Do NOT do deep content extraction. This skill builds the sheet-to-page map, nothing more. Content analysis happens on-demand for specific sheets.
• Title block = ground truth. The sheet index is a guide, but the title block on each page is the definitive source for which sheet that page is.
• Contractors provide PDFs. The input will always be a PDF. The PNGs are intermediate artifacts for analysis.
• Cache the manifest. Once built, the manifest is reusable for the entire corrections response process. No need to rebuild it.
• Handle watermarks. Some plan sets have diagonal watermarks ("Study Set", "Not For Construction", "Review Copy"). These don't affect sheet identification — the title block is still readable.

Scripts

References