feat(take-notes): add Google Slides as a source

.claude-plugin/marketplace.json

@@ -11,7 +11,7 @@
     {
       "name": "take-notes",
       "description": "Turn a YouTube video, web article, arXiv paper, or GitHub repo into didactic study notes as a self-contained HTML page, kept in a browsable local archive.",
-      "version": "1.0.1",
+      "version": "1.1.0",
       "author": {
         "name": "Daniel Verdu",
         "url": "https://github.com/davertor"

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "take-notes",
-  "version": "1.0.1",
+  "version": "1.1.0",
   "description": "Turn a YouTube video, web article, arXiv paper, or GitHub repo into didactic study notes — executive summary, the one takeaway, key points, and a timestamped or sectioned outline — as a self-contained HTML page you keep, browse, and export.",
   "author": {
     "name": "Daniel Verdu",

.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "take-notes",
-  "version": "1.0.1",
+  "version": "1.1.0",
   "description": "Turn a YouTube video, web article, arXiv paper, or GitHub repo into didactic study notes — executive summary, the one takeaway, key points, and a timestamped or sectioned outline — as a self-contained HTML page you keep, browse, and export.",
   "author": {
     "name": "Daniel Verdu",

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -42,7 +42,7 @@ body:
       description: Helps localise the failure fast.
       value: |
         ```
-        for s in render notes gallery export transcript tags; do
+        for s in render notes gallery export transcript tags retag slides; do
           uv run skills/take-notes/scripts/$s.py --selftest
         done
         ```

.github/PULL_REQUEST_TEMPLATE.md

@@ -6,7 +6,7 @@
 
 ## Checks
 
-- [ ] `for s in render notes gallery export transcript tags; do uv run skills/take-notes/scripts/$s.py --selftest; done` passes
+- [ ] `for s in render notes gallery export transcript tags retag slides; do uv run skills/take-notes/scripts/$s.py --selftest; done` passes
 - [ ] No new third-party dependency (see CONTRIBUTING.md)
 - [ ] English throughout — code, comments, commit messages
 

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
       # the whole test suite. See CONTRIBUTING.md.
       - name: Run script self-checks
         run: |
-          for s in render notes gallery export transcript tags retag; do
+          for s in render notes gallery export transcript tags retag slides; do
             echo "--- $s"
             uv run skills/take-notes/scripts/$s.py --selftest
           done

CHANGELOG.md

@@ -5,6 +5,38 @@ All notable changes to this project are documented here, in the
 release adds a section here, bumps `skills/take-notes/SKILL.md`'s `version`,
 and tags — see [CONTRIBUTING.md](CONTRIBUTING.md).
 
+## [1.1.0]
+
+### Added
+
+- **Google Slides as a source.** `/take-notes <presentation URL>` now reads a
+  deck: per-slide text, speaker notes, and a rendered image for every slide.
+  New `references/slides.md` guide and `scripts/slides.py`, stdlib only — Google
+  exports any link-visible deck as `.pptx` without an API key, and a pptx is a
+  zip of XML.
+- **A deck's diagrams reach the note.** Each slide's Google page id, read out of
+  the shape names on its notes page, addresses a 960×540 render at
+  `export/png?pageid=<id>`. Rendering the whole slide is what catches a diagram
+  drawn from native shapes — boxes and arrows that exist in no image file, and
+  that extracting embedded media would miss entirely. `slides.py` reports per
+  slide how many images, shapes and connectors it carries, which is the
+  shortlist of what is worth a figure.
+- `slides.py` says in its header whether speaker notes exist (`on 12 of 16
+  slides` / `none written on any slide`), so a deck nobody annotated cannot be
+  mistaken for notes that failed to parse.
+
+### Changed
+
+- SKILL.md's *Article figures* rule is now *Source figures*, covering articles,
+  papers and decks alike, with the same cap of 3 and the same
+  confirm-it-is-really-an-image check.
+
+### Fixed
+
+- The figure check in `references/web.md` now passes `-L`. Without it a host
+  that answers a redirect (a CDN, or Google's own image export) reads as a
+  failure, and a perfectly good figure gets dropped.
+
 ## [1.0.1]
 
 ### Fixed

CONTRIBUTING.md

@@ -20,7 +20,7 @@ There is no test suite. Each script with non-trivial logic carries its own
 asserts behind `--selftest`, and all of them must pass before a PR:
 
 ```sh
-for s in render notes gallery export transcript tags retag; do
+for s in render notes gallery export transcript tags retag slides; do
   uv run skills/take-notes/scripts/$s.py --selftest
 done
 ```
@@ -62,8 +62,9 @@ adding one guide and one routing row — never a second copy of the writing
 rules, which would immediately drift from the first.
 
 The routing table in `SKILL.md` Step 1 is matched **top to bottom, first match
-wins**. `arxiv.org` and `github.com` are `http(s)` pages, so they must stay
-above the catch-all `web.md` row or they will never be reached.
+wins**. `arxiv.org`, `docs.google.com/presentation` and `github.com` are
+`http(s)` pages, so they must stay above the catch-all `web.md` row or they will
+never be reached.
 
 ### 3. The note templates are a machine-readable contract
 

README.md

@@ -6,15 +6,15 @@
 </p>
 
 <p align="center">
-  <img src="https://img.shields.io/badge/version-1.0.0-ab2f19?style=flat-square&labelColor=191511" alt="Version 1.0.0">
+  <img src="https://img.shields.io/badge/version-1.1.0-ab2f19?style=flat-square&labelColor=191511" alt="Version 1.1.0">
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-57503f?style=flat-square&labelColor=191511" alt="MIT license"></a>
   <img src="https://img.shields.io/badge/agents-any%20Agent%20Skills%20host-57503f?style=flat-square&labelColor=191511" alt="Compatible with any Agent Skills host">
   <img src="https://img.shields.io/badge/output-html%20%C2%B7%20md%20%C2%B7%20anki-57503f?style=flat-square&labelColor=191511" alt="Output: HTML, Markdown, Anki">
 </p>
 
 **Notes you can actually learn from — not a transcript dump, not a one-paragraph summary.**
 
-Point it at a video, an article, a paper, or a repo. You get a self-contained
+Point it at a video, an article, a paper, a slide deck, or a repo. You get a self-contained
 HTML page — executive summary, the one takeaway, key points, and a timestamped
 or sectioned outline — written to `~/take-notes/html_reports/` and opened in
 your browser. They pile up into a browsable archive you own, on your disk, in
@@ -79,6 +79,7 @@ Point it at a URL. That is the whole thing:
 /take-notes https://www.youtube.com/watch?v=NiKtZgImBdY
 /take-notes https://simonwillison.net/2025/Jan/11/phi-4-bug-fixes/
 /take-notes https://arxiv.org/abs/2407.09141
+/take-notes https://docs.google.com/presentation/d/1hcGZ4U9TjZZzcGNbH2K6wYD45qwZTyo_gosCQsnHlnc/edit
 /take-notes https://github.com/ggml-org/llama.cpp
 ```
 
@@ -97,6 +98,7 @@ conversation, in any tool it's installed in.
 | **Local media** | video or audio already on disk |
 | **Web articles** | blog posts, docs pages, news articles |
 | **arXiv papers** | full text via arXiv's HTML rendering, not just the abstract |
+| **Google Slides** | slide text plus the speaker notes, and the deck's diagrams as figures |
 | **GitHub repos** | an orientation note: what it does, how it's laid out, what to read first |
 
 ### Focus
@@ -278,7 +280,8 @@ folder instead of your user directories. Update with `npx skills update take-not
 | **Whisper API key**                  | optional, only for videos without captions — Groq or OpenAI, read from `~/.config/watch/.env` |
 
 
-Web articles need none of the above — that path uses the agent's fetch tool.
+Web articles need none of the above — that path uses the agent's fetch tool —
+and Google Slides needs only `uv`: the deck reader is standard library, no API key.
 
 Both install paths land on the same `/take-notes` — neither namespaces nor
 renames it.

skills/take-notes/SKILL.md

@@ -5,7 +5,7 @@ license: MIT
 compatibility: Requires uv. Video sources also need yt-dlp and ffmpeg, plus network access; an optional Groq or OpenAI key enables Whisper for videos without captions.
 metadata:
   author: davertor
-  version: "1.0.1"
+  version: "1.1.0"
 # Claude Code extensions below — not in the agentskills.io spec, and read at the
 # top level rather than under `metadata`, which is where Claude Code looks.
 # `allowed-tools` stays comma-separated: the spec asks for spaces but marks the
@@ -104,13 +104,14 @@ regenerated, so it costs the listing and the model's choices, nothing more.
 Pick **one** reference by looking at the source, Read it, and follow it. Only the
 acquisition differs; everything after Step 2 is identical for every source.
 
-Match **top to bottom and stop at the first row that fits** — arXiv and GitHub
-links are `http(s)` pages too, so the catch-all row would swallow them.
+Match **top to bottom and stop at the first row that fits** — arXiv, Slides and
+GitHub links are `http(s)` pages too, so the catch-all row would swallow them.
 
 | Source | Read |
 |---|---|
 | YouTube URL, any other video URL yt-dlp supports, or a local media file | `references/youtube.md` |
 | `arxiv.org` (or an `ar5iv` / arXiv DOI link) — a paper | `references/arxiv.md` |
+| `docs.google.com/presentation/...` — a slide deck | `references/slides.md` |
 | `github.com/<owner>/<repo>` — a repository root, not a file, PR, or issue | `references/github.md` |
 | Any other `http(s)` page — blog post, docs page, news article | `references/web.md` |
 
@@ -123,7 +124,7 @@ Each guide hands back the same thing, and nothing more:
   date for a paper, latest release for a repo
 - **canonical URL** (plus the YouTube video ID when there is one)
 - **body** — the timestamped transcript, the article text, the paper full text,
-  or the README plus the repo's structure
+  the deck's slides and speaker notes, or the README plus the repo's structure
 
 Video sources also hand back, when yt-dlp reports them: **channel URL**,
 **published** date, **views**, a **thumbnail** URL, and the **caption language**.
@@ -302,14 +303,17 @@ Optional — include only when the source actually earns it, never as an empty h
 - `<h2>Going deeper</h2>` — what the source leaves open: unanswered questions,
   claims made without evidence, and the concrete next thing to read or try.
 
-**Article figures** — `references/web.md` returns any diagrams, charts, or
-screenshots the source page had. Include one only when it is load-bearing —
-the diagram *is* the explanation, the chart *is* the evidence — never a
-decorative photo, a header banner, or an author headshot. Cap at 3, the same
-"more than that is a dump" discipline as Key Points. Not a section of its
-own: place `<figure><img src="<url>" alt="<alt text>"><figcaption>caption</figcaption></figure>`
+**Source figures** — `web.md` and `arxiv.md` return the diagrams, charts, and
+screenshots the page carried; `slides.md` returns an image URL for every slide.
+Include one only when it is load-bearing — the diagram *is* the explanation, the
+chart *is* the evidence — never a decorative photo, a header banner, an author
+headshot, or (for a deck) a slide that is just bullets you already wrote out.
+Cap at 3, the same "more than that is a dump" discipline as Key Points. Not a
+section of its own: place
+`<figure><img src="<url>" alt="<alt text>"><figcaption>caption</figcaption></figure>`
 inline, in whichever section it supports — most often *How it works*, *Key
-points*, or *Concepts*.
+points*, or *Concepts*. Each guide says how to confirm the URL really serves an
+image before you embed it; a broken-image icon teaches nothing.
 
 ## Rules
 
@@ -329,7 +333,7 @@ points*, or *Concepts*.
   in Step 2; the structure doesn't change. Pass the matching `--lang` (`en` or `es`)
   to the renderer.
 - **Keep the HTML plain:** headings, paragraphs, lists, `<strong>`, `<em>`, links,
-  `<pre><code>`, `<blockquote>`, simple tables, and (articles only)
+  `<pre><code>`, `<blockquote>`, simple tables, and (every source but video)
   `<figure><img><figcaption>` for a source figure. No inline `style` attributes, no
   `<script>`, no classes — the stylesheet already handles presentation, and a note
   that fights it will look wrong in dark mode.

skills/take-notes/references/slides.md

@@ -0,0 +1,113 @@
+# Acquisition — Google Slides presentation
+
+Acquisition only. Return the fields listed in SKILL.md Step 1 and go back there
+to write the notes.
+
+`SKILL_DIR` is already resolved in SKILL.md — reuse it.
+
+## Read the deck
+
+```bash
+uv run "${SKILL_DIR}/scripts/slides.py" "<presentation URL>"
+```
+
+Prints the deck title, the slide count, and one `### Slide N` block per slide
+carrying its text, its **speaker notes**, its **page id**, and a count of what
+is drawn on it. No API key and no dependency: Google exports any link-visible
+deck as `.pptx`, and a pptx is a zip of XML.
+
+## Read the speaker notes — they are half the source
+
+A slide is a prompt for a person who is about to talk over it. The bullets are
+what the audience *sees*; the notes are what the presenter was going to *say* —
+the mechanism behind the diagram, the caveat behind the number, the reason this
+slide follows the last one. Notes taken from the bullets alone reproduce a deck's
+shorthand, which is the exact failure SKILL.md's "didactic means explaining, not
+compressing" rule is about.
+
+So treat the notes as body text, equal in standing to the slide's own words, and
+prefer them wherever the two disagree in depth. In practice they are where most
+of *How it works*, *Concepts*, and *Going deeper* come from — a bullet says
+"clean the traces", the note says why deterministic redaction runs before the
+LLM review.
+
+The header line says how many slides carry them:
+
+- `on 12 of 16 slides` — read every one before writing anything.
+- `none written on any slide` — a fact about the deck, not a parsing failure;
+  the script checked. Say so in *Going deeper* when the slides are thin enough
+  that the missing narration is why the notes are shorter than the topic
+  deserves.
+
+## The figures are the point of this source
+
+A deck's diagrams are frequently the entire argument, and most of them exist in
+no image file: a pipeline drawn as boxes and arrows is native Slides shapes, so
+extracting embedded media would return the screenshots and miss the diagrams.
+Rendering the whole slide catches both.
+
+Every slide has a stable image URL, and the script prints the page id for each:
+
+```
+https://docs.google.com/presentation/d/<DECK_ID>/export/png?pageid=<PAGE_ID>
+```
+
+960×540 PNG, no auth, straight into `<img src>`.
+
+The per-slide counts are the shortlist. `connectors` is the strongest signal —
+boxes alone are a layout, boxes joined by arrows are an explanation. A high
+`shapes` count is a built-up diagram or a table; `images` is a pasted
+screenshot or chart. `text only` is a bullet slide, and never a figure.
+
+Verify before embedding, the same check `web.md` uses:
+
+```bash
+curl -sIL -o /dev/null -w "%{http_code} %{content_type}" "<png URL>"
+```
+
+`-L` is not optional here: the export answers `307 application/binary` and
+redirects to the rendered image, so without it every figure looks like a failure.
+
+Anything but `200 image/png` means the deck is not link-readable — drop the
+figure rather than shipping a broken image, and fold the caption into the prose.
+
+Two things to be honest about in the note when it matters: the URL renders the
+deck's **current** slide, so a figure drifts if the deck is later edited, and it
+breaks entirely if the deck's sharing is tightened.
+
+Cap and placement follow SKILL.md's **Source figures** rule — at most 3, inline
+in the section each one supports, never a gallery.
+
+## Map the output to the Step 1 fields
+
+| Step 1 field | From |
+|---|---|
+| title | `**Title:**` — the deck's own name |
+| byline | the presenter or team named on the title slide; the script cannot report one, so read it off slide 1, and fall back to `Google Slides` when the deck names nobody |
+| span | `**Slides:**`, written out in the note's own language — `16 slides` or `16 diapositivas` |
+| canonical URL | `**Deck URL:**` — the bare `/edit` form, without the `#slide=` fragment the user's link carried |
+| body | the `## Slides` blocks, text and speaker notes together |
+
+A deck has no video ID, so Step 5 renders the article layout. Its `Section
+outline` is the deck's own arc: group the slides into the 6–15 moves the talk
+actually makes, and link each to its slide with the deep link the script printed
+(`.../edit#slide=id.<PAGE_ID>`). One line per slide is a table of contents, not
+an outline.
+
+A **focus** narrows which slides matter, not which are fetched — the whole deck
+arrives in one download either way.
+
+## Failure
+
+The script exits non-zero and says which of these it hit:
+
+- **not link-readable**: the export answers with a sign-in page. Say so and stop.
+  The fix is the user's: set sharing to *Anyone with the link → Viewer*, or paste
+  the content. Do not write notes from the title.
+- **not a Slides URL**: a Docs or Sheets link is a different product, and a
+  `/presentation/` link that is really a published `/pub` page is a web page —
+  use `references/web.md`.
+
+A deck that is mostly screenshots with three words per slide has no body to take
+notes from. Say that rather than padding one out; the speaker notes are the last
+place to check before giving up.

skills/take-notes/references/web.md

@@ -45,7 +45,8 @@ image URLs when asked, in document order, with captions attached. It returns
 none of that unasked, so don't drop this line.
 
 **Before embedding one as a `<figure>`, confirm it actually serves an image:**
-`curl -sI -o /dev/null -w "%{http_code} %{content_type}"` on the URL. A page
+`curl -sIL -o /dev/null -w "%{http_code} %{content_type}"` on the URL — with
+`-L`, since a CDN that answers a redirect is not a failure. A page
 behind bot protection (Cloudflare and similar) returns an HTML challenge page
 instead of image bytes for any out-of-browser request — and critically, no
 header fixes this, because it also blocks the `<img>` tag in the finished