Use when creating illustrated PPT / infographic slide decks that tell a continuous story about a topic (tech sharing, onboarding, concept walkthrough, lesson recap, 萌系分享图, 故事 slide). Triggered by "做一套 PPT 讲 xxx", "生成故事图讲 xxx", "用萌系/日漫风做 xxx 分享图", "做一套故事 slide", or direct /ppt-anything invocation.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Use when creating illustrated PPT / infographic slide decks that tell a continuous story about a topic (tech sharing, onboarding, concept walkthrough, lesson recap, 萌系分享图, 故事 slide). Triggered by "做一套 PPT 讲 xxx", "生成故事图讲 xxx", "用萌系/日漫风做 xxx 分享图", "做一套故事 slide", or direct /ppt-anything invocation.
ppt-anything
Generate a cohesive series of anime-chibi-style illustrated infographic slides that tell a story about a user-provided topic. Default aesthetic: SD chibi Japanese anime characters + clean infographic layout + pinpoint-accurate Chinese text. (Users can override the style explicitly; otherwise this is what they get.)
Core principle — content is the hero
The content is the main subject. Characters, poses, scenes, and decorations exist to make the content come alive — not to steal attention from it.
Read hierarchy per slide: 内容清晰可读 > 人物表达情绪 > 装饰点缀氛围. If a pose or decoration makes the viewer read slower, it's a bug. Whenever a choice competes — e.g., a bigger character portrait vs. a larger title — the content wins.
Everything downstream (character selection, reference-image discipline, layout rotation, pose vocabulary, decoration density) is a supporting mechanism. When one of those mechanisms pulls focus from the copy or the beat, it's malfunctioning — cut or shrink it.
Runtime layout (must know)
This skill is self-contained — no external sub-skills required.
Single image client. Default: Google Gemini official. Also supports Gemini-compatible bridges and OpenAI-compatible bridges. Run --help for provider options.
Batch parallel image client. Reads JSON manifest, all slides generated simultaneously, failed slides auto-retry. Used by fast mode.
~/.ppt-anything/providers/<name>.toml
API credentials ([auth].api_key + [auth].base_url + [auth].docs_url + [auth].auth_style). Source of truth for keys. The agent never reads this file directly — generate-image.py does.
~/.ppt-anything/characters/<slug>/<slug>.{md,png}
Character library. Skill reads from here at runtime.
~/.ppt-anything/styles/<style-name>/
Style packs. Default anime-chibi-default/ ships with the skill.
~/.ppt-anything/demo/<YYYY-MM-DD>-<HHMM>-<slug>/
Auto-archive of every generated deck (outline + slides + html + meta). Includes time to prevent same-day collisions.
Single-file HTML packager (base64, only for WeChat/email use cases).
Provider iron rules (read before first run)
Agent does NOT read provider toml files. The toml stores api_key. Reading it = pulling the key into the agent's conversation context (cached, persisted in memory systems, replayed across turns, leaked between agents). To check whether config is ready, call generate-image.py and let it fail with a clear status — its error messages tell you which field is missing without exposing the key. Never cat / Read any *.toml file body.ls of the filenames is fine; contents are not.
Handing api_key / base_url to the agent is a risk operation. When the user's path is "I'll edit the toml myself", the agent sees zero key bytes — safe. When the user's path is "agent, please write the toml for me", the agent must explicitly warn ("your key will enter my context, the recommended path is option A") and require a second confirmation before touching the key string.
Agents that register a new provider MUST verify connectivity. Writing the toml is not enough — call tools/register-provider.py which uses the cheapest model + smallest prompt to actually round-trip one image, and only marks verified=true when that succeeds. A failed verify means the toml goes back to verified=false and the user is told what to check.
First-run wizard (no provider configured yet)
Trigger (mandatory, BEFORE asking the user any briefing question): The agent runs ~/.claude/skills/ppt-anything/tools/check-providers.py. The script reads provider tomls internally and outputs a key-free summary on stdout:
Exit 0 + READY: <names> → at least one provider is configured. Note the names; surface them in Step 2 as the choices for "which provider".
Exit 1 + NEED-CONFIG: ... → STOP. Open the conversation below before doing anything else.
The agent does NOT cat / Read any toml. check-providers.py does the reading; agent only sees ok/incomplete summaries.
"I notice ppt-anything has no usable provider configured yet. Two paths:
Path 1 — Google official (recommended)
The shipped ~/.ppt-anything/providers/google.toml already has base_url / docs_url / auth_style filled with the official values. You only need to:
Open google.toml in your editor and paste it into [auth].api_key
Strongly recommended. I never touch your key bytes.
Path 2 — Third-party bridge (risky)
If you've subscribed to some third-party Gemini-shape or OpenAI-image-compatible bridge, I can write the toml for you. You'd give me:
base_url (gateway URL)
api_key (giving this to me means it enters my context — caching / memory / multi-agent transcript leakage are all on the table)
docs_url (I'll WebFetch it to confirm request/response shape)
auth_style (google for X-Goog-Api-Key, bearer for OpenAI-style Authorization)
Reconfirm: if you go Path 2, I will see your key. If you'd rather I never see it, switch to Path 1 — cp google.toml my-bridge.toml, edit it yourself.
Which path? (1 / 2 / something I haven't listed)"
If the user picks Path 2, the agent runs the connectivity self-test via tools/register-provider.py (which writes the toml AND verifies it round-trips one image before returning). See § Setup tool below.
generate-image.py reads credentials from ~/.ppt-anything/providers/<name>.toml itself. The agent does not need to involve itself — call the script and let it handle the secret.
Setup tool — register-provider.py
The agent uses this when registering ANY non-default provider. It:
Writes ~/.ppt-anything/providers/<name>.toml from the user-supplied fields.
Generates one tiny image via gemini-2.5-flash-image (~$0.04, ~10s) to verify base_url + api_key + auth_style actually work end-to-end.
Marks [provider].verified = true only on success. On failure, the toml is left with verified = false and the script exits non-zero so the agent flags the user to check the offending field.
Invocation pattern (agent-side; key never goes on the command line — pipe via stdin):
If the user prefers Path 1 (Google official, fill the key themselves), the agent does NOT run register-provider.py — it tells them to edit google.toml directly, then they re-invoke the skill and the agent verifies on first generation.
When to use
User asks for a story-telling illustrated slide deck / 故事 PPT / 萌系分享图 / 绘本 slide
Topic needs multiple slides to walk through (onboarding, tech talk, concept recap, workflow explanation)
User wants anime-chibi aesthetic combined with Chinese text
When NOT to use
Single standalone picture with no slide context → use generate-image directly
Only a cover image needed → generate-image with one prompt is faster
Workflow
Provider readiness gate (run BEFORE everything else, including library snapshot). Run ~/.claude/skills/ppt-anything/tools/check-providers.py. Read the stdout summary:
Exit 0 → at least one provider is READY or verified. Remember the list; you'll surface it as Step 2's provider-choice options. Proceed to Step 1.
Exit 1 → no provider usable. DO NOT proceed to briefing. Surface the problem to the user and run the first-run wizard (§ First-run wizard above): walk them through Path 1 (self-fill google.toml) or Path 2 (tools/register-provider.py for a bridge). Resume Step 1 only after the user has finished setup AND a re-run of check-providers.py exits 0.
Anti-pattern (do NOT do this): don't proceed to briefing/outline/generation when the gate fails — you'll waste user time + collect a brief that can't be executed because the rail isn't ready. The gate is the FIRST thing in this skill, not an afterthought triggered by a 401 mid-flow.
Library snapshot, briefing second. Before asking the user any open-ended brief questions (audience / register / slide count / character preference / etc.), do these reads in parallel:
ls ~/.ppt-anything/characters/ — list available character slugs
ls ~/.ppt-anything/styles/ — list available style packs
ls ~/.ppt-anything/providers/ — list provider toml files (filenames only; do NOT cat them — agent never reads provider toml)
For each character slug, Read the character's <slug>.md profile to know what it looks like (you'll need this to judge fit). For each style, glance at <style>/manifest.md if it exists.
Then open the briefing with the user — lead with what's actually in the library, not a generic 5-question form. Example shape:
Anti-pattern (do NOT do this): Send a 5-question form (视角 / 切片 / 基调 / 角色 / 张数) without first surfacing what's available. The library constrains feasibility; ignoring it makes briefing rounds blow up.
Topic-character fit judgment: if the user's topic obviously needs a kind of character the library can't serve (e.g. user wants a 写实校园怀旧, library only has abstract OC water droplets), proactively flag that mismatch in the briefing — offer two paths: (a) accept abstract symbolic handling with the existing characters, (b) add a new character now (Branch B/C of step 2). Don't proceed with a topic that can't work and discover it 3 rounds later.
Quality gates (self-check every run)
T1: Reused a character from the library without re-searching (when the character was already on disk)
T2: New character got a profile written BEFORE generation, not after
T2b: Every character in the deck has a local reference image on disk (~/.ppt-anything/characters/<slug>/<slug>.<ext>) AND Claude has Read it (vision input) at least once this session. No text-only character goes into generation.
T3: Full outline approved by user before the first generate.py call
T4: EVERY generate.py call passes the character reference image(s) as --ref; slides 2+ ALSO pass the previous slide as --ref. No slide relies on daisy-chained previous slides alone.
T5: Each slide uses a DIFFERENT layout from the previous slide (no 5× identical ribbon-banner template)
T6: Each slide uses a DIFFERENT character pose from the previous slide
T7: 生成模式被显式问过 — 高质量(顺序+每张看图复盘) vs 快速(batch-generate.py 全量并行+失败自动重试)。没有静默默认。高质量模式下每张生成完 AI 都 Read 过该图; 快速模式下通过 batch-generate.py 全部同时发出, 失败自动重试, 不中途 Read。
T11: provider 选定后, model + 分辨率都用了 provider 范围内的顶配。具体参数参考 generate-image.py --help。降配只在用户明示"省钱/草稿/不敏感"时触发。
T8: Every element on every slide earns its place — no filler character / decoration / sentence added to meet a quota. If asked "why is this here?" you can answer with a story reason, not a rule reason.
T9: Design intent was articulated per slide and shown to the user BEFORE any generation — covering beat, layout choice, pose choice, copy, decorations (with reasoning)
T10 (content-first sanity check): On each finished slide, the viewer's eye lands on the title/copy first, then the character, then the decorations. If a character or decoration is stealing the scene, shrink or reposition — characters serve the content, not the other way around.
Failing any gate = restart from the failing step. Do not patch on top of a broken gate.
Style rules
See ~/.ppt-anything/styles/<active-style>/style_guide.md. Default aesthetic: SD chibi 2.5-head-body characters, hand-painted watercolor finish, clean pastel infographic layout, LXGW WenKai Chinese font, content-first hierarchy. Key sub-sections to consult:
§ 3 Character pose vocabulary — pick a fresh pose each slide; avoid mid-action poses and abstract props (§ 3 Pose traps)
§ 4 Layout variation — pick a different layout each slide; match layout to beat shape (§ 4 Layout × beat fit); for memory / emotion / lived-experience beats, consider a custom scene over A–H (§ 4 Scene-ify)
§ 6 Background — clean, restrained, no "magical forest" filler
§ 7 Decoration density — context decides, not a number; flow/timeline slides = near-zero extras
§ 12 No emoji, ever (hard rule) — zero Unicode emoji in any text slot; describe marks as painted motifs
§ 13 Chinese glyph clarity (soft hint) — keep lines short, split long phrases across two lines
Deviate only when the user explicitly requests a different style. Drift within a single deck is the #1 failure mode.
Prompt template
See ~/.ppt-anything/styles/<active-style>/prompt_template.md. All slots are <<ALL_CAPS>>. Never ship to generate.py with a slot still in place.
Character library
See ~/.ppt-anything/characters/README.md for profile schema, the mandatory reference-image rule, and the two branches (public vs user-provided). Every new character — whether researched online or given to you by the user — must be persisted as BOTH a profile .md AND a local image file. Text description without a verified image = character drift in production.
User pays for the wrong rail / wrong quality / wrong key
If ls ~/.ppt-anything/providers/*.toml shows >1 entry, ASK the user which to use — do not silently default
Same layout on every slide
Monotonous deck, reader disengages
Vary layouts per ~/.ppt-anything/styles/<active-style>/style_guide.md § Layout variation (T5)
Same character pose every slide (e.g. "holding notebook")
Dead deck, no motion
Pick fresh pose from vocabulary per slide (T6)
Padding a slide with filler (extra characters / decorations / copy) to "look rich"
Dilutes the main message; filler ≠ richness
Every element earns its place or gets dropped (T8)
Character in the visual center with content pushed to the side
Character steals focus from the message
Characters serve the story — place content at the visual anchor, characters frame/support
Style-library miss judgment (mirror of the character rule above): if the user's prompt names a style direction the library doesn't have (e.g. user says "科技风" / "复古风" / "商务风" / "赛博朋克" but styles/ only has anime-chibi-default), the correct response is to BUILD a new style pack — not to relay confusion back as "我对你的话有 N 种理解". Library miss IS the signal to build, not a license to ask "which of my interpretations did you mean?". Lead with the build framing:
"库里现在只有 [既有 style], 没有 [用户说的 style] — 我去建一个新 pack。在 [用户说的 style] 内部你拍方向: A 硬路线 / B 折中 / C 跟既有 style 混搭 + 该方向 motif。"
Then once the user picks A/B/C, build the new pack (manifest.md + style_guide.md + prompt_template.md + outline_template.md mirroring anime-chibi-default/'s structure) and proceed. Building a new style pack is a normal workflow branch, not an exception path. Anti-pattern: responding with "我对你说的 X 风格有 2-3 种理解，是哪种？" — that puts the inferential burden back on the user; they already named the style, the missing piece is direction within it.
Prompt research when building a new pack or using a new model: Before writing prompt_template.md, WebSearch for high-quality prompt patterns specific to the target model (e.g. "gpt-image-2 best prompt structure", "gemini image generation prompt tips"). Different models respond to fundamentally different prompt shapes: diffusion models want tag-soup, instruction-tuned models want editorial briefs, some models need negative prompts, others ignore them. The built-in anime-chibi-default/ ships with model-specific variants (prompt_template.md for Gemini/Seedream, prompt_template_gpt_image.md for gpt-image-2) precisely because one template does not fit all. When building a new style pack, research the model first, then write the template. When the user's config.toml points to a model the existing pack hasn't been tested with, search for that model's prompt best practices before generating.
Clarify what's still missing — including provider choice and generation mode. After the library snapshot you know what's available, BUT YOU STILL ASK USER A FEW DECISIONS BEFORE TOUCHING generate-image.py:
Topic-side: register (serious/playful) if ambiguous from topic, slide count if user didn't say, any new-character/new-style needs that the snapshot revealed.
Provider choice — MANDATORY when more than one provider toml exists. Do NOT silently default. If ls ~/.ppt-anything/providers/ shows multiple configured .toml files, explicitly ask: "providers/ 里有 [list]，这次用哪个？". Only auto-select when there's exactly ONE configured provider on disk. (Reason: the user pays per call, may want different model/quality, may have one provider keyed and one not — choosing for them costs them money on the wrong rail.)
Resolve characters — and ALWAYS anchor with a real reference image. Text description alone is never enough; the model will drift toward its training-data bias (e.g. a 2012 version of a character that's been redesigned in 2024). Every character used in the deck must end this step with BOTH (a) a profile file ~/.ppt-anything/characters/<slug>/<slug>.md AND (b) a local reference image ~/.ppt-anything/characters/<slug>/<slug>.<ext> that Claude has Read (vision input) at least once this session.
Three branches:
Branch A — character already in library: check ~/.ppt-anything/characters/<slug>/<slug>.md. If the profile exists AND ~/.ppt-anything/characters/<slug>/<slug>.<ext> exists on disk, Read both. Done.
Branch B — public character, not yet in library: WebSearch + curl the canonical image into ~/.ppt-anything/characters/<slug>/<slug>.<ext> → Read it with vision → write the profile file from WHAT YOU SAW, not from text summaries. If multiple canonical versions exist (old vs redesign), confirm with the user which version to use, and record the choice in version_note.
Branch C — user-provided image (character can't be found online / OC / obscure): ask the user for the image path/URL if not given → copy into ~/.ppt-anything/characters/<slug>/<slug>.<ext> (never leave in /tmp) → Read it → write the profile with image_provenance: user-provided, leaning more on explicit feature description since the model has no latent memory to summon.
See ~/.ppt-anything/characters/README.md for the full schema, two-branch workflow details, and the real failure this rule was written to prevent. Skipping the Read step is the #1 source of character drift. Do not skip it even if the name feels famous.
Plan the story arc — think, don't fill a form. Decide slide count based on what the topic needs. For each slide, you are designing a beat in a story — think through it, don't fill a checklist:
What is the ONE core message of this slide? What emotion should the viewer feel?
Which layout (A–H from ~/.ppt-anything/styles/<active-style>/style_guide.md § Layout variation) best fits that message — OR does this beat earn a custom scene (scrapbook, corkboard, chat window, recipe card, desk workspace...) per § Scene-ify? Scene-ify for memory / emotion / lived-experience; A–H for teaching / reference / taxonomy. Before committing: run § Scene-ify's modifier check (POV / tense / register / simultaneity). If a vocabulary row fits 1:1 after the check, trust it — don't drift to look creative. If no row holds your modifiers, invent.
How much copy does this beat need? A big emotional moment might be three bold words; a teaching beat may need a paragraph. Let the content drive density, not a rule.
Which characters serve this beat, in what pose? Pick poses that amplify the emotion, not to fill a slot.
Which decorations reinforce the beat without stealing focus? Could be 2, could be 10 — context decides. Less is often more when the copy is already information-rich; more welcome when the copy is sparse and the story leans visual.
Hard constraint (not a quantity rule, a variety rule): adjacent slides MUST use different layouts AND different character poses. This is about rhythm, not quota.
Anti-rule: never pad a slide with decorations / characters / copy just because some template said to. If an element doesn't earn its place, drop it.
Use ~/.ppt-anything/styles/<active-style>/outline_template.md as a thinking scaffold — articulate your design intent in your own words, not by filling blanks.
Gate: share the outline, wait for approval. Write out your design intent for every slide (using ~/.ppt-anything/styles/<active-style>/outline_template.md as a thinking scaffold) and show it to the user. Include enough so the user can tell: is each slide purposeful? Do layouts and poses vary? Does any slide feel over/under-filled? Do NOT call generate.py yet. Wait for explicit approval.
Generate slides — execute the mode the user picked in Step 2, not a default. Before filling prompts:
Default — lightweight external-reference HTML (recommended for almost every deck):
Compress each PNG to visually-lossless WebP (q=95; typically ~80-90% smaller than PNG for watercolor illustrations) and wrap them in a ~3 KB HTML that references them via relative paths. Opens instantly; folder is self-contained and droppable into GH Pages or any static host.
Defaults: q=95 WebP, source PNGs removed after compression, output at <dir>/index.html. Flags:
--lossless — bit-perfect WebP (~30% smaller than PNG; use only if user explicitly requires no lossy compression)
--keep-png — preserve the source PNGs (default deletes them to keep the folder minimal)
--quality N — override q=95 (q=90 for more aggressive, q=98 if you worry about fine text)
Requires cwebp (on macOS: brew install webp).
Alternative — self-contained single-file HTML (rare):
Use only when the user needs the deck as ONE file (WeChat/email attachment, air-gapped share, can't deploy a folder). Base64 inflates file size ~33% on top of PNG — a 5-slide deck becomes ~70 MB HTML that takes several seconds to open every time.
Open locally:open <dir>/index.html (macOS) / xdg-open (linux). Controls: ← → space for nav, F for fullscreen, edge-click for prev/next. Don't skip this step — a silent "here's the file path" delivery makes the user do work.
Optional — GH Pages prep. ASK the user: "Want a GitHub-Pages-ready folder for this deck?" Only prepare on explicit yes. For the external-reference variant the deck folder IS the deploy folder — just drop a .nojekyll file in it. For the base64 variant, create <slug>-deploy/ with the html copied in as index.html + .nojekyll. Never push, never gh repo create, never touch their GitHub account on their behalf.
Report: HTML path + total deliverable size (print output of the build script already gives this) + one-paragraph story recap.
Copy-pasting the previous slide's design with different words
Deck feels flat and machine-made
Each slide is a distinct beat — think it through, don't template it
Researching a character but skipping the profile write
Defeats the library's purpose
Write first, then generate
Writing a character profile from a text description WITHOUT reading an actual image
Model drifts to the training-data default version (e.g. 2012 老版 when you wanted 2024 redesign); by slide 3 the character is unrecognizable
Branch A/C: always download or receive a real image, Read it as vision input, then write the profile from what you saw. Gate T2b.
Passing only the previous slide as --ref for slides 2+ (daisy-chain)
Character features drift a little each hop; by slide 5 the accumulated drift is visible
Every slide passes the ORIGINAL ~/.ppt-anything/characters/<slug>/<slug>.<ext> as --ref in addition to the previous slide. Gate T4.
User says "this character can't be found online, here's a picture" and you ignore the picture, reasoning from their verbal description
Their words miss 3/5 visual details that matter for consistent regeneration (eye shape, prop texture, exact hair part)
Branch C: copy the provided image to ~/.ppt-anything/characters/<slug>/<slug>.<ext>, Read it, let the image drive the profile.
Passing vague English gloss for Chinese text
Model hallucinates characters
Pass exact Chinese string in quotes, demand accuracy explicitly
Asking model to "be creative with font"
Chinese glyphs break
Always specify LXGW WenKai (霞鹜文楷) style
Picking Layout D for a mindset-flip beat (e.g. 破除共识 / 用户≠上帝)
Didactic layout neutralizes the punch — message reads like a manual page
Use Layout G (split) or B (hero) — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 4 Layout × beat fit
Defaulting to Layout D + bullet list for an emotional / memory / lived-experience beat (e.g. 回忆录 / 情怀 / 复盘 / "这周我们经历了…")
Turns a story moment into a manual recap — the feeling doesn't land; user reads bullets, not the memory
Consider a custom scene (scrapbook, corkboard, chat window, recipe card...) where content lives inside the metaphor and characters act it out — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 4 Scene-ify
Grabbing the nearest § Scene-ify vocabulary row without checking the beat's hidden modifiers (POV / tense / register / simultaneity)
Container holds the topic but not the shape — the slide "looks right" but the emotion doesn't land
Run the modifier check first — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 4 Before picking — ask the beat's hidden modifiers
Departing from a § Scene-ify vocabulary row that was already 1:1 right, to "avoid looking table-dependent" or "show creativity"
Overcorrection — the easy answer was correct; you picked a worse scene to perform thinking
If the modifier check comes back clean, trust the row — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 4 If a row fits 1:1 after the modifier check, trust it
Writing UI-design words (icon, card, drop-shadow, badge) into a watercolor deck
Triggers flat-vector Material Design style, clashes with the rest of the deck
Use watercolor vocabulary — see ~/.ppt-anything/styles/<active-style>/prompt_template.md top-level rule 6
Dynamic mid-action chibi poses ("about to press stamp", "mid-swing") + abstract digital props (.md file, code file)
Model freezes mid-action as awkward, and abstract props render as generic rectangles
Use terminal-state poses + concrete iconic substitutes (gift box, scroll, target) — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 3 Pose traps
Outline shows per-slide design but buries the slide ORDER inside prose
User spots order problems only after generation — reframing finished beats is expensive
Separate a glanceable one-line order sequence in the outline — see ~/.ppt-anything/styles/<active-style>/outline_template.mdBefore showing the user
Emoji render as flat OS glyphs, photobomb the watercolor deck
Zero emoji in text slots. Use color-coded pills + plain text, or describe the motif as "a watercolor brush-stroke tick mark" — see ~/.ppt-anything/styles/<active-style>/style_guide.md § 12
Delivering a base64-embedded HTML (build-html-ppt.py) by default
3k PNGs base64-encoded inflate ~33%; a 2-slide deck is ~38 MB, a 5-slide deck is ~70+ MB. Takes several seconds to open every time, feels broken, gets closed. User in this project has explicitly flagged this as "太重打开很慢".
Default to tools/build-html-ppt-external.py (external WebP refs, q=95, ~3 KB HTML + 2-3 MB per slide). Only use the base64 version when the user explicitly needs a single-file deliverable (WeChat/email attachment, air-gapped share).
Keeping 3k PNGs alongside the WebPs in the final deck folder
Doubles folder size for no reason — HTML references .webp, PNGs are dead weight
build-html-ppt-external.py deletes source PNGs by default after successful compression. Only pass --keep-png if the user asked to keep originals.
Running build-html-ppt-external.py without cwebp installed
Script exits with "cwebp not found" and the user has to install + re-run
Before first use on a fresh machine: brew install webp (macOS) / apt install webp (debian). Verify with command -v cwebp.
