The AI compares both pages as images — and keeps fixing until they match.
Run a formatted report through an ordinary translator and you get
text. The charts, the score bars, the boxed findings — the parts that
actually carry the result — are gone. Page-for-Page renders the original and
the translation side by side, looks at them, and corrects whatever does
not line up — then looks again. Same trim size, same charts with the
same values, same colours, same running heads.
It is an open-source Claude skill. Install it and
run your own documents through it.
Built in a day, then put through its own review loop —
three language directions, every layout value measured rather than estimated,
and the bugs that surfaced folded back in. MIT licensed.
SourceOutputSourceOutput
Trim size
A4 · 8.27 × 11.70 in
Chart bands
40 / 20 / 40 %
Band fill
#F2D79A
Bar fill
#7A2E6D
Row pitch
0.547 in
Axis scale
70.5 px / unit
Values read out of the
source file itself — not estimated.
How it works
Measure, rebuild, then check by eye.
Every colour, band width, row pitch and
axis position already exists in the source file. The work is reading them out
instead of approximating them by eye — which is the difference between a page
that matches and one that only nearly does.
1
Measure the original
Trim size, exact hex fills, chart geometry and type scale are read from the
file at 150 dpi. Values that exist only as pixels — a normative range,
a mean marker — are recovered by measurement rather than dropped or invented.
2
Pick the route that holds
Where the layout is dense and graphical, the text layer is replaced in place
and every logo and chart stays the original vector. Where the target language
runs longer and needs to reflow, the page is rebuilt from the measurements. The
fit is checked before committing either way.
3
Compare, correct, repeat
Both files are rendered to page-pair images and read side by side: bar
directions, arrow directions, label attachment, missing glyphs, and every number
against the original. Whatever is off gets fixed and the pages are rendered
again. Two or three rounds is normal — and none of it is visible to a text
diff, which is why the check has to be a look.
Proof
Diagrams are where a translation breaks.
Prose is the easy part. A flow chart is
not: the meaning lives in arrow direction, in which label is attached to which node,
and in the order the nodes are read. Every word can be translated correctly and the
picture can still say something else — and a text diff will report no
difference at all. That is the reason the check has to be a look.
One-way cycle, four nodes, a label under each and a caption pointer
into the middle. Same positions, same directions, same attachment — the
artwork is the original vector, only the text layer changed.The first pass drew one arrow of the cycle backwards. Nothing in the
text was wrong, so nothing flagged it — it was caught by rendering both
pages and reading them, then redrawn. This is that failure reproduced on the
public sample; two or three such rounds is normal.
Then every page, side by side
A five-page assessment report,
Korean to English. Source on the left of each pair, output on the right. The graded
band chart keeps its bar values and sides, the facet chart keeps its midzone marks,
and the practitioner chart keeps its range bars, mean ticks and scores.
SourceOutput
Pages 3–4 — the style clarity chart and the facet chart with
its in-preference / midzone / out-of-preference bands.
SourceOutput
Pages 1, 2 and 5 — cover, the eight-cell definitions grid, and the
practitioner summary with normative range bars.
About this sample
The document shown is a fictional
assessment instrument — invented publisher, invented respondent, invented
scores — built so the example can be shown publicly. It does not correspond
to any real instrument, organisation or person. The reversed arrow above is a real
failure class, reproduced here on that sample rather than screenshotted from the
document it was first caught on.
Run it yourself
Install it and point it at your own PDF.
Two commands. Then ask Claude to translate a PDF and
mention that the format should stay the same — it picks the route, measures
the source, lifts the logos, and checks its own output. Everything below is in the
repository: the skill, the scripts, the reference notes, and the sample document
shown above.
# in Claude Code or Cowork
/plugin marketplace add modocai/pdf-translate-layout
/plugin install pdf-translate@modoc
Needs poppler-utils,
ImageMagick, and Python pillow · pymupdf · playwright. A bundled
preflight script checks all of it.
What you get
A measurement toolkit — trim size, palette,
band edges, row pitch, axis scale, and chart values recovered from pixels.
Two translation routes, with the fit checked
before either is committed.
Artwork extraction, so logos are the original
marks rather than hand-drawn approximations.
Page-pair comparison tooling for the review step.
Tested reference notes for CJK, right-to-left,
Thai, Indic and European targets.
No tokens, no setup
Or just email it to us.
Running this costs Claude tokens, and not everyone
has them to spend — or wants to install a toolchain to translate one
document. Email the PDF and the language you want and we will run it and send the
result back.
No account, no payment page. We are doing this
because we want to see what people bring us.
Tell us the target
language. If the document is confidential, say so before you send anything and
we'll agree how to handle it — or just run the skill yourself, which keeps
the file on your own machine.
Honest scope
Verified on Korean → English, English → Korean and
English → German. Other pairs use the same method; right-to-left
targets are documented but not yet verified end to end.
Where a logo cannot be lifted from the source it is approximated, and that is
stated in the delivery. Output is a translation, never a reissue of an original.
Translations that run shorter than the source can leave whitespace the
original does not have. We report it rather than padding with invented text.
Scanned PDFs with no text layer need an OCR pass first.
