Drawing Diff

Overview

Take two revisions of the same sheet or set, find every place they differ, and write a plain list of what changed to docs/drawing-diffs/. This replaces the Bluebeam Compare Documents pass an office runs on every consultant re-issue, and the slower version of it: printing both revisions and holding them up to a window.

Two engines run, and they see different things. Text extraction catches dimensions, notes, tags, room names, ratings, and title-block edits, and reports them as exact values. A rendered overlay catches linework that moved when no text changed: a wall shifted, a door swing flipped, a hatch pattern replaced. Neither one alone is a complete answer, and the report always says which ran.

Scope and limits — read first

What this does. Pairs sheets by the sheet number printed in the title block, diffs each pair two ways, and reports added, removed, and edited text with its value on both sides, plus a rectangle for every region of linework that changed. Every changed region is shown to you as a before/after crop.

What this does not do.

• It does not tell you a change is safe. It reports that a dimension went from 3'-6" to 4'-0". Whether that breaks the stair, the clearance, or the structure is code-compliance-review, design-stair, or a person.
• It does not decide which revision is newer. You name the old file and the new file. The skill does not read dates and reorder them.
• It never reports "no changes" from one engine. If text extraction found nothing and the overlay did not run, the answer is "text found no changes and geometry was not checked", never "nothing changed".
• It does not compare two sheets of different shape. A half-size print against a full-size sheet is fine, because that is one ratio applied to everything and both engines take it out. A different aspect ratio is a different sheet, and the run stops rather than pretending.
• It does not read a scanned sheet's text. If either side has no extractable text, that side is a scan or fully flattened, and the text engine is skipped for the pair. Comparing a vector sheet against a scan of it would report every line as deleted. The overlay still runs, and the report records which side was unreadable.
• It does not trust revision clouds. Clouds are compared against detected changes, in both directions. It does not use them as the list of changes.
• It does not install anything. Missing PyMuPDF stops the run and prints the install command for you.

What you need before starting. Two PDFs of the same sheet or set, and PyMuPDF importable. Comparing a full set is one run per sheet pair, so sets past about 40 matched sheets are worth narrowing to the disciplines you care about, and the skill asks before starting one.

When to use

Trigger on:

• Two pasted PDF paths of the same sheet or set, with any comparing verb.
• Explicit asks: "what changed", "compare these revisions", "diff these drawings", "check this against the last issue", "did they change anything in the re-issue".
• A consultant re-issue, bulletin, ASI, or addendum arriving against a set you already have.

Do NOT trigger on:

• One PDF with markups on it. That is a redline to answer, not a revision to compare: use rfi-writer or drawing-reader.
• Two different sheets (A-101 against A-102). There is nothing to pair.
• Spec book revisions. Text-only documents diff better with doc-reader.
• A set that just needs bookmarking or a sheet index: drawing-set-bookmarker.

Workflow

Step 1 — Bootstrap if needed

Check whether docs/drawing-diffs/ exists with an index.md. If not, run:

python scripts/init_workspace.py

The script creates docs/drawing-diffs/ from assets/drawing-diffs-index-template.md and verifies PyMuPDF is importable. If it is missing the script prints the install hint (pip install PyMuPDF) and exits non-zero — surface this to the user and stop. Do NOT pip-install on the user's behalf.

If the project has no docs/ folder at all, confirm with the user before bootstrapping.

Step 2 — Confirm which file is which

Ask the user which PDF is the older revision if it is not stated. Do not infer it from filename, modified date, or page count. Getting this backwards inverts every "added" and "removed" in the report and nothing downstream catches it.

Step 3 — Match sheets

Run:

python scripts/match_sheets.py \
  --old "<old.pdf>" --new "<new.pdf>" \
  --out "docs/drawing-diffs/.cache/<base>-pairs.json"

The script reads each page's title-block corner, pulls the sheet number by pattern, and pairs the two files on that number. It reports four groups: matched pairs, sheets only in the old file (removed), sheets only in the new file (added), and pages whose sheet number could not be read.

For pages in that last group, read them with the Read tool (pages: "<n>") and identify the sheet number by eye, then re-run with --override <page>=<sheet-number> (repeatable). Never pair by page order to close the gap. Sets gain and lose sheets between revisions, and page order stops meaning anything the first time that happens.

Report the four counts to the user before comparing. Added and removed sheets are findings on their own, and often the biggest ones.

Step 4 — Compare each matched pair

For each pair:

python scripts/compare_sheet.py \
  --old "<old.pdf>" --old-page <n> \
  --new "<new.pdf>" --new-page <m> \
  --sheet "<A-101>" \
  --out-dir "docs/drawing-diffs/.cache/<base>"

The script writes <sheet>.json plus one before/after PNG pair per changed region. What it returns:

• text.added / text.removed / text.edited — each with the string and its position on the sheet. edited means text at the same spot reads differently, which is where dimension and rating changes land.
• regions — rectangles where the rendered pages differ, largest first, with the crop paths.
• alignment — the pixel offset applied before overlaying, the residual, and the paper-size ratio. A re-plot shifts the whole sheet a few pixels and a half-size print scales it; both are taken out before comparing. If alignment.residual_pct is above about 40, treat the region list as unusable and say so: the sheets did not align, and what follows is not a list of changes.
• skipped — which engine did not run, and why. no-text names the side that is a scan; aspect-mismatch means the two sheets are different shapes.

text.sheet_shift_pt and text.sheet_scale record what was normalised out before diffing. Non-default values are worth a line in the report: they mean the sheet itself was re-plotted or re-sized, which is a fact about the issue even when no content changed.

Default caps are 10 regions per sheet and 1% of a tile's pixels to call it changed. Raise --max-regions when a sheet comes back saturated, and say in the report that the list was capped.

Step 5 — Read the changed regions

For each region crop pair, read both PNGs with the Read tool and describe the change in the terms the drawing uses: "partition at grid C/4 moved about 600mm east", "door 108 swing reversed", "roof hatch changed from tapered insulation to ballast". Name a grid reference when the crop shows one.

The script says where. This step says what. Do not paraphrase the script's rectangle coordinates into the report as if they were the finding.

If a crop is ambiguous, say so and give the coordinates. Do not invent a reading of a region you cannot see clearly.

Step 6 — Check the revision clouds both ways

Read the new sheet's revision block with the Read tool and record what the sheet claims changed. Then compare against Step 4 and 5:

• Changed, not clouded. A detected change with no cloud around it. This is the finding that pays for the whole run. List these first.
• Clouded, not changed. A cloud with nothing detected inside it, usually a cloud left over from the previous revision that nobody cleared.
• Revision block silent. Changes detected but the block lists no new revision entry.

Never treat the cloud set as the change set. Consultants miss clouds, and the sheets where they missed are exactly the sheets worth this run.

Step 7 — Write the report

Copy assets/diff-report-template.md and fill it out to:

docs/drawing-diffs/<old-rev>-to-<new-rev>-<date>.md

One section per sheet, sheets with changes first. Findings inside a sheet are ordered: unclouded changes, then text edits, then geometry, then clouded but unchanged. Every finding carries the sheet number, the location, and both values where there are two. End with the methods note from references/diff_methods.md stating what ran and what that could not see.

Step 8 — Update the project index

Run:

python scripts/append_to_index.py \
  --old "<old.pdf basename>" --new "<new.pdf basename>" \
  --sheets-compared <N> --sheets-changed <M> \
  --report "<report filename>"

Appends one row to docs/drawing-diffs/index.md. Do NOT edit index.md directly with Edit — table alignment breaks.

Step 9 — Report

Tell the user:

• Path to the report.
• Sheets compared, sheets changed, sheets added, sheets removed.
• The unclouded changes, by sheet. These go first, in the chat, not just the file.
• Any sheet where an engine was skipped, and which one.
• Any sheet that failed to align.

Rules

• The user names which file is older. Never infer it.
• Sheets pair on the sheet number read from the title block, never on page order, filename, or position in the set.
• Text values are quoted verbatim on both sides. 3'-6" to 4'-0", not "the dimension grew".
• Every reported region is backed by a crop the user can open. No finding without an image or an extracted string behind it.
• A skipped engine is stated in the report, on the sheet where it was skipped. Silence reads as "checked and clean".
• Revision clouds are evidence, compared in both directions, never the source of the change list.
• Output lives in docs/drawing-diffs/ of the user's project, never in the skill folder or /tmp. Crops in .cache/ are working files, not deliverables.

Anti-patterns

• Pairing sheets by page index because the sheet number would not read. Ask, or use --override.
• Reporting "no changes" after a text-only run. Say which engine ran.
• Listing the revision block's claims as the change list. That is the thing being checked, not the answer.
• Feeding both full sheets to vision and asking what changed. Dense sheets defeat it and the misses are silent. The scripts narrow it down first.
• Reporting a text diff against a scanned sheet. One side with no text means the text engine is off for that pair, and the report says which side.
• Aligning two sheets of different shape "close enough". Stop and say so.
• Editing docs/drawing-diffs/index.md with the Edit tool. Use append_to_index.py.
• Auto-installing PyMuPDF for the user. Surface the command and stop.
• Judging whether a change is acceptable. Report it and hand off to the code and design skills.

Resources

• scripts/init_workspace.py — bootstrap docs/drawing-diffs/ and verify PyMuPDF
• scripts/match_sheets.py — pair pages between two PDFs by title-block sheet number
• scripts/compare_sheet.py — text diff plus aligned raster overlay for one sheet pair, with region crops
• scripts/append_to_index.py — append a comparison row to docs/drawing-diffs/index.md
• references/diff_methods.md — what each engine sees and misses, and the methods note to paste into every report
• references/_provenance.md — sheet-numbering source and verification date
• assets/diff-report-template.md — per-comparison report skeleton
• assets/drawing-diffs-index-template.md — initial markdown table for the project-wide comparison log