Schedule Extractor

Overview

Pull tabular drawing-set schedules (doors, windows, room finishes, equipment) out of a PDF as a normalized CSV plus a markdown preview, and log them to a per-project schedule index in docs/schedules/. This replaces the hand-typing pass that happens every time a consultant ships a CD set without the live Revit schedules attached.

When to use

Trigger on:

• Pasted path to a drawing-set PDF together with a schedule type ("extract the door schedule from this set", "pull the finish schedule").
• Explicit asks: "schedule to CSV", "tabulate the door schedule", "give me a finish schedule spreadsheet".
• Cross-checking values against a separate consultant submission ("compare the architect's door schedule to the contractor's submittal").

Do NOT trigger on:

• Bookmark/index requests for the entire set — use drawing-set-bookmarker instead.
• Drawing-set completeness review — use drawing-checklist.
• Marked-up/redlined schedules needing field-question handling — use rfi-writer.
• General questions about what a finish schedule should contain — answer directly without invoking this skill.

Workflow

Step 1 — Bootstrap if needed

Check whether docs/schedules/index.md exists. If not, run:

python scripts/init_workspace.py

The script creates docs/schedules/ from assets/schedules-index-template.md. No external Python deps are required for this skill — extraction uses the Read tool directly. If the project has no docs/ folder at all, confirm with the user before bootstrapping.

Step 2 — Identify schedule type and locate pages

Determine which schedule(s) the user wants:

• door — keyed off Mark / Type / W / H / Frame / HW Set columns.
• window — keyed off Mark / Type / W / H / Glazing / Operation columns.
• finish — room finish schedule, keyed off Room # / Floor / Base / Wall / Ceiling.
• equipment — keyed off Tag / Description / Mfr / Model.

If the schedule type is ambiguous, ask the user before proceeding.

Locate the schedule pages either by:

• Page range supplied by the user ("schedule is on sheets A-601 to A-603").
• Sheet index from drawing-set-bookmarker if available (docs/drawing-sets/<base>-index.csv filtered to Discipline=A and SheetType=6).
• Direct probe: read sheet titles via the Read tool, looking for the literal phrase "DOOR SCHEDULE" / "WINDOW SCHEDULE" / "ROOM FINISH SCHEDULE" / "EQUIPMENT SCHEDULE" in the title block.

Step 3 — Vision read of schedule pages

Read the schedule pages with the Read tool, chunking 5–10 pages per call. For each page:

1. Capture the table header row verbatim (column names exactly as printed).
2. Capture every body row as a list of cells, in column order.
3. Note any merged cells, sub-headers, or continuation indicators ("CONT'D ON A-602").
4. Note any footnotes / "Notes:" blocks below the table — these often contain hardware abbreviations or fire-rating overrides referenced by row.

Persist the raw extraction as JSON at docs/schedules/.cache/<schedule-type>-raw.json:

{
  "schedule_type": "door",
  "source_pdf": "<basename>.pdf",
  "pages": [
    {
      "page_index": 142,
      "sheet_number": "A-601",
      "headers": ["Mark", "Type", "Width", "Height", "Frame", "HW Set", "Fire", "Notes"],
      "rows": [
        ["101A", "F1", "3'-0\"", "7'-0\"", "HM", "HW-1", "20 min", ""],
        ...
      ],
      "footnotes": ["HW-1: lockset, closer, hinges (3)"]
    }
  ]
}

If a row is illegible or has a missing cell, write the cell value as "" (empty string) and add confidence: low to that row entry. Never invent values.

Step 4 — Normalize columns to standard schema

Run:

python scripts/normalize_schedule.py \
  --raw "docs/schedules/.cache/<schedule-type>-raw.json" \
  --type <door|window|finish|equipment> \
  --out "docs/schedules/<basename>-<schedule-type>-schedule.csv"

The script maps captured header names to the standard schema using a header-synonym table. Columns it cannot map are appended at the end with their captured names preserved. The script never silently drops columns.

Door, window, and finish each have a schema reference in references/, listing the column meanings and the type codes behind them. Equipment has no reference file; its columns are the inline list below and the header row in assets/equipment-schedule-template.csv.

Standard schemas (full detail in references):

• Door: Mark, Type, Width, Height, Thickness, Material, Frame, HardwareSet, FireRating, Remarks
• Window: Mark, Type, Width, Height, Material, Glazing, Operation, Remarks
• Finish (room): RoomNumber, RoomName, Floor, Base, NorthWall, EastWall, SouthWall, WestWall, Ceiling, Remarks
• Equipment: Tag, Description, Manufacturer, Model, Power, Remarks

For dimensional fields, preserve the imperial format as printed (3'-0", 7'-0"). Do NOT silently convert to decimal feet or metric.

Step 5 — Write a markdown preview

Beside the CSV, write a markdown table preview at docs/schedules/<basename>-<schedule-type>-schedule.md. The preview shows the first 20 rows for quick visual review before the user opens the CSV. Use the same column order as the CSV.

Step 6 — Update the project index

Run:

python scripts/append_to_index.py \
  --date 2026-05-08 \
  --source "<basename>.pdf" \
  --type <schedule-type> \
  --row-count <N> \
  --csv "<basename>-<schedule-type>-schedule.csv" \
  --preview "<basename>-<schedule-type>-schedule.md"

Appends a row to docs/schedules/index.md. Do NOT edit index.md directly with Edit — table alignment breaks.

Step 7 — Report

Tell the user:

• Path to the CSV and the markdown preview.
• Row count, broken down by confidence: low rows that need manual verification (with their Mark / Tag).
• Any captured columns that did not map to the standard schema (preserved at the end of the CSV).
• Cross-reference flags: hardware sets referenced but not defined in the table footnotes; fire ratings inconsistent with NCS sheet-type series; missing Mark values.

Rules

• Cell values are quoted verbatim from the PDF — never normalize dimensions, never reformat fractions, never expand abbreviations on output.
• Every row visible in the table must produce a row in the CSV. Page row count in equals output row count out.
• Standard column schema is per references/<type>_schedule_schema.md. Unknown columns are preserved at the end with original names.
• Output files live in docs/schedules/ of the user's project, never in the skill folder or /tmp.
• Vision-extracted rows with confidence: low are reported to the user, not silently merged.
• Cache JSON in docs/schedules/.cache/ is treated as resumable state, not deliverable — do not surface it as an artifact.

Anti-patterns

• Filling in missing cells from "common sense" (e.g. assuming a 3'-0" door if Width is illegible). Mark confidence: low and leave the cell empty.
• Converting dimensions ("3'-0"" → 36" → 0.9144 m"). The schedule is the contract drawing — keep verbatim.
• Expanding hardware-set abbreviations ("HW-1" → "lockset, closer, hinges (3)") in the cell. The footnote stays in references/hardware_sets.md guidance, but the cell value stays "HW-1".
• Editing docs/schedules/index.md with the Edit tool. Use append_to_index.py.
• Skipping the markdown preview step. The user reviews the preview first; the CSV is downstream.
• Treating tabula-py / camelot output as authoritative without vision verification — those PDF table extractors silently misalign cells on dense schedules. Vision read is the source of truth here.

Resources

• scripts/init_workspace.py — bootstrap docs/schedules/
• scripts/normalize_schedule.py — map raw extracted JSON to standard CSV per schedule type
• scripts/append_to_index.py — append a processed-schedule row to docs/schedules/index.md
• references/door_schedule_schema.md — door schedule columns + type codes (NCS / industry)
• references/window_schedule_schema.md — window schedule columns
• references/finish_schedule_schema.md — room finish schedule columns (CSI format)
• references/hardware_sets.md — common HW-set abbreviation glossary (project-specific caveat)
• assets/door-schedule-template.csv — header row for door CSV
• assets/window-schedule-template.csv — header row for window CSV
• assets/finish-schedule-template.csv — header row for room finish CSV
• assets/equipment-schedule-template.csv — header row for equipment CSV
• assets/schedules-index-template.md — initial markdown table for the project-wide schedules index