Imported from dddpaul/claude-skills (plugins/presentation/skills/pptx-compare/SKILL.md). Install upstream with npx skills add dddpaul/claude-skills --skill pptx-compare. Copyright stays with the author.

PPTX Deck Comparison

Three scripts that measure the gap between two decks. They are diagnostic and opinion-free: they parse any two .pptx files and consider neither of them correct. Nothing here encodes a house style — for that, see the sibling pptx-arch-style skill.

When to apply

1. Aligning a generator against a manual reference. Someone hand-built a deck in PowerPoint and you are writing code to reproduce it. Run the structural comparison after every generator change to see what still differs, and dump_slide.py on the reference to read the exact numbers to encode.
2. Regression-checking a rebuild. Compare the previously committed .pptx against a freshly built one. Anything the comparison reports that you did not intend to change is a regression.
3. Reading exact values off a foreign slide. dump_slide.py prints one slide as frames, fills, lines, wrap and text by run, in EMU and inches — faster and more exact than clicking through the PowerPoint inspector.

Boundary: this skill measures, it does not fix

The comparison reports the gap between two decks. It does not edit the generator, rewrite a deck, or propose a patch — closing the loop is the caller's job. The scripts read both files and write only their report and, with --render, their images. Expect to run compare → change the generator → rebuild → compare again, several times; the skill is the measuring instrument in that loop, not the loop.

Skill independence

The plugin's skills are mutually independent; the consumer does the wiring. pptx-compare must not call lint.py or postprocess-effectlst.py from pptx-arch-style, and pptx-arch-style must not call anything here. A build pipeline that wants generate → post-process → lint → compare assembles those four steps itself.

The temptation is to have the comparison run the style check too, "for completeness". Resist it: that would tie the two skills' versions together and stop the plugin from updating them piecemeal. The two differ in nature — pptx-arch-style is normative (references/rules.yaml plus a conformance check, with a verdict), pptx-compare is diagnostic (two decks in, a list of differences out, no verdict).

Scripts

scripts/compare_decks.py — structural comparison

uv run scripts/compare_decks.py REF.pptx GEN.pptx [--pos-tol INCHES]
                                [--fold-engine-artefacts]
                                [--render] [--dpi N] [--outdir DIR] [--report FILE]

For every slide it captures each shape — text by run, font face, size, weight/style, colour, fill, outline, alignment, indents, and the x/y/w/h frame in EMU — matches the shapes of the two decks by a pair cost (text similarity, geometric distance, shape kind), and prints the discrepancies per slide. Shapes that match nothing on the other side are reported as only in ref / only in gen rather than force-fitted into a misleading diff.

Runs are compared after merging adjacent runs of identical formatting, walked side by side by the characters each covers, so a paragraph the two engines split in different places reports nothing. The run[N] in a finding numbers those merged spans, so it need not line up with the run[N] dump_slide.py prints for either deck.

• --pos-tol INCHES — coordinate slack; x/y/w/h deltas at or below it are not reported. Default 0.040in. A different generator pair needs a different value: widen it until engine rounding stops showing, then leave it there so real drift still surfaces.
• --fold-engine-artefacts — drop the findings the tool itself attributes to the engine pair, and count only what is left. Off by default; see "Folding engine artefacts" below.
• --render — additionally rasterise both decks (see below).
• --report FILE — write the same text to a file as well as stdout.

Exit codes: 0 decks match within tolerance, 1 at least one discrepancy, 2 not comparable (slide counts differ, or a render failed).

scripts/dump_slide.py — read one slide

uv run scripts/dump_slide.py DECK.pptx N

N is 1-based, matching PowerPoint's own numbering. Prints frame, fill, line, wrap and text by run for every shape, each dimension in both EMU and inches. It imports compare_decks as a plain module and shares its shape model, so what you read here is exactly what the comparison compares.

scripts/pixel_diff.py — compare renders

uv run scripts/pixel_diff.py REF_DIR GEN_DIR --outdir DIR
                             [--threshold N] [--max-diff FRACTION]
                             [--zoom N:LEFT,TOP,RIGHT,BOTTOM]

Takes two folders of PNGs, reports the fraction of differing pixels per page, writes an overlay with the differing pixels highlighted in red, and tiles those overlays into one contact sheet. --zoom N:LEFT,TOP,RIGHT,BOTTOM crops page N of ref, gen and overlay to a region for a close look.

This script knows nothing about pptx — it works on any two folders of identically-sized images. Feed it the output of compare_decks.py --render, which puts both decks through one shared dpi so the pages line up.

External dependencies

The structural comparison and the slide dump need only python-pptx — no LibreOffice, no Node. If soffice or pdftoppm is missing, --render fails with a clear message and the structural comparison still runs.

Sandbox caveat

--render shells out to LibreOffice, which writes a user profile and temporary files. Under a sandbox that only permits writes inside the repository, point --outdir at a directory inside the repository (and git-ignore it) or disable the sandbox for that command. --outdir has no default beside the script — the skill lives in a read-only plugin cache — and falls back to ./_compare_out under the current working directory.

Reading the output

Not every reported discrepancy is a defect. When the two decks come from different engines — a PowerPoint-authored reference against a pptxgenjs-generated deck — some differences are structural artefacts of the engine pair and will never go away, no matter how the generator is written. Before chasing a discrepancy, check it against references/engine-differences.md, which lists the four known ones.

Folding engine artefacts

--fold-engine-artefacts drops the findings the tool has already attributed to the engine pair and counts only what is left, so a deck differing solely by those reports Total: 0 and exits 0.

The default is unchanged: without the flag every difference is still listed, engine artefacts included. The flag is opt-in and subtracts nothing else. A run-count mismatch is folded only once both paragraphs coalesce to the same formatted text — merging adjacent runs is lossless only when they carried identical formatting, so a split whose runs differ in font, size, weight or colour is drift the engine pair does not explain and survives the fold. The runs themselves are compared after that same merge, walked side by side by the characters each covers rather than paired by index, so where the two decks split a paragraph has no bearing on what is reported.

Which mode to use depends on what you are reading the output for:

The convergence loop is the case the flag exists for. Aligning a generator against a hand-built deck (use case 1 above) means running the comparison after every generator change; once the only remaining differences are engine artefacts, an unfolded report still prints several dozen lines and "done" has to be judged by eye, line by line. Folded, the same run prints zero and the loop has a stopping condition a script can test:

uv run scripts/compare_decks.py ref.pptx gen.pptx --fold-engine-artefacts && echo converged

A folded report says so in its header, so a filtered report is never mistaken for a full one. Fold to decide whether you are done; drop the flag to read what is actually left.

Known limitation: groups are opaque

A grouped shape is compared as one shape; its contents are not inspected. The parse walks the top-level shape tree of each slide, so a p:grpSp contributes its own name and frame and nothing else — two decks whose groups hold entirely different text can still be reported as OK. Grouping is idiomatic in hand-built PowerPoint decks, which is exactly use case 1 above, so do not read a clean structural report as proof that two decks match until you know neither contains groups.

Two ways round it: ungroup in the reference before comparing, or fall back on --render plus pixel_diff.py, which sees the rendered result and so is blind to how shapes are nested.

Typical loop

uv run scripts/compare_decks.py ref.pptx gen.pptx --report compare.md
# fix the generator, rebuild gen.pptx, repeat until only known engine artefacts remain

uv run scripts/compare_decks.py ref.pptx gen.pptx --fold-engine-artefacts
# same comparison, artefacts folded: "Total: 0" and exit 0 mean converged

uv run scripts/compare_decks.py ref.pptx gen.pptx --render --dpi 144 --outdir ./_cmp
uv run scripts/pixel_diff.py ./_cmp/ref ./_cmp/gen --outdir ./_cmp/diff
# open ./_cmp/diff/contact-sheet.png, then --zoom into whatever looks wrong

Fixture decks under scripts/tests/fixtures/ are regenerable with uv run scripts/tests/gen_fixtures.py.