sheet-review — export → evidence → verdict
A fit-to-screen look cannot see a 0.3 pt stray line on a 36×24 sheet. In one studio project, "PDF reviewed" was claimed while stray bed linework and a colored legend over uncolored symbols went through, and that miss is logged 30 times. This skill makes the review cover every tile, and it keeps what a script proved separate from what the visual pass saw.
Scripts live in this skill's scripts/ folder; run commands from the skill folder.
Order
- Pin the artifact. Record the PDF path and its INDD, with the sha256 and mtime of both. Review the exported PDF, not a description of it.
- InDesign audit (when the INDD is open). It only reads and never launches InDesign.
Limit: both routes below need an
execute_extend_scripttool from a fork of the adb-mcp Adobe connector. That tool is not in upstream adb-mcp yet. If it isn't available, say the audit was not run and list it under not checked. The PDF preflight (step 3) does not need it.- MCP: read
indesign_audit.jsx, setTARGET_DOC_NAMEto the document name, and pass the text to an InDesign MCP tool namedexecute_extend_script(see the repository README). Passapp_namewhen both the release and Beta InDesign are running. - Without MCP (Codex, shell):
<adb-mcp venv>/bin/python scripts/run_indesign_audit.py --doc NAME.indd --out <checks>/indesign-audit.jsonIt audits whichever InDesign is running. Add--app "<name>"only when the release and the Beta are both running; the error lists the running names. Don't write a year from memory: Creative Cloud replaces the apps on upgrade (2026 became 2027 in November 2026). - If
unsavedChangesis true, the PDF does not represent the live layout. Say so.
- MCP: read
- PDF preflight:
Dependencies are PyMuPDF, numpy, Pillow and scipy (all inpython3 scripts/pdf_preflight.py SHEET.pdf --out <checks>/sheet-review-<date> \ --expect-size 36x24 --palette <project>/standards/linework.md --compare PREVIOUS.pdfrequirements-arch-tools.txt). Exit 1 means a FAIL. Outputs arereport.md,report.json,p<N>-overview.pngwith the tile grid labeled,tiles/p<N>-r<row>c<col>.pngat 150 dpi (--tile-dpi 300for fine linework) andp<N>-diff.png. Always pass--comparewhen a previous export exists: the changed-region list is the cheapest way to catch a stray line or an accidental move. - Visual pass on every tile, in order. For each tile write
cleanor the issue with its tile id. Look at the overview first for composition at fit-page distance. Then, per tile:- Stray linework: construction lines, duplicates, lines crossing furniture, beds or text, lines outside a drawing frame, broken hatch boundaries.
- Legend ↔ symbol: every legend swatch appears on the drawing it keys, and every
colored category on the drawing is in the legend. A colored legend over uncolored
symbols is a fail.
swatch-only-colorfindings point at suspects. - Line hierarchy: cut/poché heaviest, then medium, then light, per the project's linework standard.
- Scale: each stated scale (the preflight lists every notation) matches the placed size. A placed drawing that isn't at 100% in the audit makes its stated scale wrong unless the label was changed to match. Scale bars and north arrows are where the standard asks for them. NTS appears only where intended.
- Text: readable at print size, nothing clipped, no placeholders, consistent titles.
- Images: pixelation or banding. Section cut marks match the section titles.
- Report in three columns: verified by script / seen in visual pass / not checked. Keep status words exact: exported ≠ printed ≠ submitted. Never call a sheet ready with an open FAIL, an unreviewed tile, or unsaved INDD changes.
Reading the automated findings
| Check | Meaning | Watch for |
|---|---|---|
page-size FAIL | wrong size or orientation | — |
font-not-embedded FAIL | the plotter may substitute | — |
low-ppi-image | effective PPI at placed size | bbox can extend past the page on cropped images |
small-text | spans below --min-text-pt (6) | sizes are listed, the smallest is located |
placeholder-text | lorem / TODO / TBD / XXX / TK | — |
hairline-strokes | strokes < 0.1 pt may vanish in print | Make2D linework weights |
swatch-only-color | a non-neutral fill used only in ≤ 6 small shapes, never stroked | raster images aren't scanned; if the keyed items are in a placed image, confirm by eye |
off-palette-fills | fills > 0.5 in² not near any palette hex (grays and near-black ignored) | pass the palette the project actually uses |
changed-region (INFO) | pixels changed vs. the previous export | every region should map to an intended edit |
INDD notAt100 | a placed drawing is scaled | the stated drawing scale must be recomputed |
Eval
tests/seeded_eval.py SHEET.pdf --work <scratch> --palette … copies a real sheet and plants
five defects: a stray line, a legend swatch with no symbols, an off-palette fill, 3 pt text
and "TBD". It then runs the preflight with --compare and scores detection against that
answer key. The last run, on a real 36x24 review sheet (2026-10-05), caught 5/5. Re-run it after
any change to the script. Stray linework that existed before the first export is invisible
to --compare, so the visual pass stays mandatory.
Design critique (conventions, argument, composition) is a separate step, not covered here. This skill is the evidence pass, and it runs before that critique.