Pptx To Slide Md

PPTX → slide-Markdown

Convert an existing .pptx back into the plain-text format the user writes slides in, so a finished deck becomes editable source again.

Target format

One block per slide. --- separates slides; === separates slide content from speaker notes. Presenter instructions (images, animations, demos) live in [brackets] inside the content area. If the deck has sections, each section begins with a [Section N: Section Name] label before the first slide in that section.

[Section 1: Introduction]

[Slide 1]

# Slide Title

Slide content in markdown: bullets, numbered lists, tables, **bold**, *italic*, ~~strikethrough~~, [links](url).

[image: assets/slide3_img1.png — "alt text"]

[animation — 2 click steps:
  click 1: entrance of "First point"
  click 2: entrance of "Second point"]

===

**Speaker Notes:**

Notes in markdown.

---

[Slide 2]

# Next Slide Title

...

===

**Speaker Notes:**

*(none)*

Each slide is preceded by a [Slide N] label (N is the slide's position in the deck, so a subset extraction keeps the original numbers). --- separates slides and === separates content from notes, as before.

Blank lines between elements ensure Markdown viewers render each block as a separate paragraph rather than collapsing consecutive lines into soft-wrapped text.

The user also uses [^] inside speaker notes to mark a click. See the note on animations below for why extraction puts animation detail in the content area rather than fabricating [^] positions in the notes.

How to run it

The work is done by scripts/extract.py (uses python-pptx, already available).

python scripts/extract.py  -o 

Options:

• -o, --output — output path (default: .md in the cwd).
• --slides — subset to extract, e.g. --slides 3-20 or --slides 1,4,7-9

(1-based; default: all). Numbers in [Slide N] stay absolute.

• --images-dir DIR — where to write extracted images (default:

_images/ next to the output file).

• --no-images — don't write image files; still emit [image: filename] inline.

After running, read the output file and make it available to the user. Surface the generated .md (and the images folder) using whatever file-sharing mechanism your runtime offers — a file preview or attachment if one exists, otherwise point the user to the output path. Don't paste a huge deck inline if it's long — write the file and show a representative excerpt.

What gets extracted

• Title → # Heading (from the slide's title placeholder).
• Body text → bullets by indent level (-, nested with two spaces),

numbered lists for auto-numbered paragraphs, plain paragraphs when a paragraph has no bullet. Run formatting becomes **bold**, *italic*, ***both***, ~~strikethrough~~, and [text](url) for hyperlinks. Adjacent runs with the same formatting are merged; soft line breaks within a paragraph become spaces.

• Tables → Markdown tables (pipes inside cells are escaped).
• Images → extracted to the images folder and referenced as

[image: — ""] (alt text included when the shape has it).

• Charts → [chart: — ""] (not the underlying data).
• SmartArt / diagrams → [SmartArt/diagram — text: ...] when text is

reachable, otherwise a note that the text isn't extractable.

• Speaker notes → Markdown under **Speaker Notes:** (*(none)* if empty).
• Animations → best-effort, see below.
• Hidden slides → flagged with [hidden slide].
• Sections → [Section N: Name] emitted before the first slide of each

section (only when the deck actually uses sections).

• Grouped shapes are walked recursively; shapes keep document order, with the

title hoisted to the top.

Animations — what is and isn't recoverable

PowerPoint stores animations as a time-node tree (`) that python-pptx` does not expose, so the script parses the raw slide XML. Be honest with the user about the boundaries:

• Reliable: the number of click-advance steps, and which shape/paragraph

each step reveals (the script maps the animated paragraph range back to its text). Build-by-paragraph reveals (one bullet per click) come through.

• Reliable: the effect category — entrance, exit, emphasis, motion path.
• Not reported: the specific effect name (Fade vs. Wipe vs. Fly-In). The

preset-ID → name mapping is version-dependent and easy to get wrong, so the script deliberately reports only the category rather than guessing.

• By design: animation detail goes in a [animation — …] block in the

content area (matching the user's convention that presenter instructions live in brackets). The script does not insert [^] click markers into the notes, because where each click belongs in the authored notes prose is not stored in the file and can't be reconstructed faithfully. If the user wants [^] markers in the notes, offer to add them using the click count as a guide.

• If timing XML is present but can't be parsed, the slide gets

[animation present — could not parse (…)] instead of failing.

Build sequences across slides (always surface these)

Decks often split one idea across several consecutive slides: a list that grows one bullet per slide, or a fixed background with different things overlaid. The script flags runs of consecutive slides that share a title and prints them to stderr under BUILD SEQUENCES DETECTED, classifying each as either an additive build or a shared base with changing overlays.

Always extract these as separate slides (one block each, as normal) — do not merge them automatically. But after extraction, tell the user which sequences were detected and offer to compact each into a single slide with the build expressed as progressive-disclosure instructions and [^] click markers in the notes. Only compact if the user agrees. This applies to additive lists, overlay-style sequences, and restyled finales (e.g. a slide that repeats the previous one with text struck through).

Optional follow-ups

• Describe images: if your runtime can view images, open the extracted image

files and, if the user wants, replace [image: file] with a short description.

• Verify a tricky deck: if the output looks off (e.g. a text box that should

be a heading came through as a bullet, or vice-versa), the slide XML can be inspected directly; bullet detection relies on explicit buNone/buChar/ buAutoNum formatting and falls back to "body placeholders bullet, everything else is plain", which is a heuristic, not a guarantee.

Edge cases worth knowing

• Embedded media, OLE objects, and equations are not extracted (only noted if

they carry text).

• Merged table cells repeat their content per the underlying model.
• A slide with no title placeholder is emitted as # (Slide N — untitled).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AssafTzurEl
• Source: AssafTzurEl/assaf-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.