Manus Slides

Manus Slides

Full-pipeline presentation tool: topic -> outline -> slides -> export (PPTX/PDF/HTML).

Three modes:

1. Whiteboard (recommended) — 24 AI styles via Gemini (gemini-3.1-flash-image-preview / Nano Banana 2) → PPTX
2. HTML — 13 styled templates via slide_templates.py, editable, Playwright screenshots → PPTX/PDF
3. Image — Gemini generates custom images per slide → PPTX

Quick Start — Whiteboard Mode

# Recommend styles for a task
python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py recommend "investor pitch deck for AI startup"

# Create config JSON with slide prompts
cat > slides.json << 'EOF'
{
  "title": "My Presentation",
  "style": "whiteboard",
  "slides": [
    {"id": "slide_01", "prompt": "TITLE in bold: \"My Topic\"\nSUBTITLE: \"Key subtitle\"\nDraw a relevant illustration below."},
    {"id": "slide_02", "prompt": "TITLE: \"Problem Statement\"\nDraw diagram showing the problem with icons and arrows."}
  ]
}
EOF

# Generate all slides
python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py generate slides.json ./output

# Output: ./output/slide_01.png, slide_02.png + presentation.pptx + preview.html

Quick Start — HTML Mode

echo '[{"title":"Title","summary":"Welcome","slide_template_key":"cerulean"}]' > outline.json
python ~/.claude/skills/manus-slides/scripts/slide_manager.py init "Title" outline.json ./project
python ~/.claude/skills/manus-slides/scripts/slide_export.py html ./project ./presentation.html

AI Styles (24 total via whiteboard_generator.py)

Manus Originals (7) — Best visual quality

Manus Hybrid (7) — стили по мотивам Manus, рендер на Gemini

Bonus Styles (10)

Manus 1.6 Image Themes (6) — стили по мотивам image-режима Manus (2026-07-02)

Наши промпт-стили, повторяющие внешний вид image-тем Manus 1.6; рендер идёт на Gemini. Образцы слайдов для сверки: references/manus-theme-samples/<Theme>__<model>/. Карта 26 тем (режимы, модели, палитры, соответствие HTML-темам): references/manus-26-themes.md.

⚠️ Manus's "Sketch" theme = our sketch-notebook (light cream ink-doodle). Our legacy sketch (white chalk on DARK paper) is a different look — kept for back-compat. Manus's "Whiteboard" = our whiteboard. ⚠️ pixel footer tags and dossier/vellum decorative stamps are auto-invented by the model if you don't specify them — put real footer/tag text in the slide prompt to avoid gibberish (standard "specify all text" rule).

Что даёт эмуляция тем Manus 1.6 и где её предел

• Наши STYLES-промпты собраны по внешнему виду: превью тем плюс 8-слайдовые образцы. Это эмуляция стиля, а не точная копия — совпадения слайд-в-слайд не будет.
• Живьём генерация в Manus прогонялась только для Sketch/nano-banana; остальные темы сверялись по образцам слайдов.
• Из 19 HTML/react-тем Manus: 6 закрываются нашими AI-image стилями (glamour, amber, arctic, neon, patina, onyx), 9 ложатся на наши HTML-шаблоны и палитры (см. references/manus-26-themes.md), под 4 (Emerald, Mist, Linen, Mahogany) нужен новый photo-based HTML-шаблон — готового соответствия нет.
• Результат «ровно как в Manus» получается только в самом Manus.

Image Model Selection (env MANUS_SLIDES_MODEL)

whiteboard_generator.py defaults to gemini-3.1-flash-image-preview (Nano Banana 2 — fast, cheap, excellent Cyrillic). Override per run for premium decks:

# premium deck with Nano Banana Pro:
MANUS_SLIDES_MODEL=nano-banana-pro-preview python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py generate slides.json ./output editorial

⚠️ gemini-3.5-flash is text-only (no image output) — do NOT set it here. It's the newest flash but generates text, not slide images. (Verified via API 2026-07-02.)

Production notes for the 5 Manus themes (validated on multi-slide decks, 2026-07-02)

The refined prompts are hardened against the #1 failure mode: the model invents gibberish in any element the style declares mandatory but the slide content leaves empty (subtitle, footer tags, "Est." line). Subtitle / footer / accent lines are now CONDITIONAL — omitted (not invented) when you don't supply them. Still, for full control, specify every piece of on-slide text in the prompt.

• editorial — auto-switches layout: title/section slides → dark photo-split; content slides (lists/diagrams) → light cream "paper" layout (matches Manus). Hint it by starting a slide prompt with TITLE SLIDE. or CONTENT SLIDE..
• pixel — give footer tags explicitly (Footer tags: "Агенты · Автономность · 2026") and a window label; if omitted they're dropped cleanly (no more garbled bar). Great for product-design / tech-nostalgia decks.
• vellum — do NOT wrap terms in quotes in the prompt (the model renders the quote marks literally on the slide). Seal chop is auto-anchored under the title / bottom-right.
• dossier — the red Est. YYYY line is an opt-in signature: add RED ITALIC ACCENT LINE: "Est. 2026" on the title slide; omit on content slides. Don't quote list lead-ins. Lists ≤5 items and 2×3 comparisons hold best; props auto-arrange around the edges.
• etching — production-ready as-is; footer running caption + roman numeral render cleanly only when you specify them.
• Reference look for all 7 Manus image themes (incl. whiteboard/sketch): references/manus-theme-samples/<Theme>__<model>/.

HTML-Only Styles (13 via slide_templates.py)

cerulean, cobalt, emerald, basalt, mist, sand, linen, alabaster, quartz, mahogany, ginkgo, sunset, lavender

Scripts

All in ~/.claude/skills/manus-slides/scripts/:

whiteboard_generator.py (AI Mode)

Generates slides using Gemini Image (gemini-3.1-flash-image-preview / Nano Banana 2) in any of 24 AI styles.

Config format:

{
  "title": "Presentation Title",
  "style": "whiteboard",
  "slides": [
    {
      "id": "slide_01",
      "prompt": "Content description — what to draw/write"
    }
  ]
}

Dependencies: pip install google-genai python-dotenv Pillow python-pptx

Post-processing exec-sketch decks — production gotchas (learned in prod)

A logo is composited AFTER generation (the prompt no longer reserves an empty corner — see the white-box fix below). Hard-won lessons when assembling/maintaining real decks:

• Logo is a separate composite, not part of the prompt. Keep a logo PNG with alpha (your brand mark) and img.alpha_composite() it post-gen. Position is a design choice — be consistent across the WHOLE deck. A safe default for client decks = top-RIGHT: right edge ≈ 0.976*W, top 0.038*H, height 0.072*H → x = W - lw - int(W*0.024); y = int(H*0.038). (Top-left works too but clips left-aligned titles — see below.)
• Keep the pre-logo RAW PNGs (exec_slides_raw/, no logo). To move the logo to the other side / resize / rebrand, re-run the composite on raws — never try to "erase" a baked-in logo.
• ⭐ WHITE-BOX ROOT CAUSE = the prompt itself. The old exec-sketch prompt said "Leave the very top-left corner clean and empty (… a logo is composited there later)" — telling Gemini to "leave a corner empty" makes it literally draw an empty white box/container there. FIXED at the prompt level (2026-06-14): the style string now says the cream bg "COMPLETELY covers all four corners — NO white rectangles, NO empty blocks, NO containers, NO reserved boxes in any corner" and no longer reserves a logo corner. New decks won't have the box. The logo is composited on top of the cream afterwards regardless.
  • For decks generated with the OLD prompt (box already baked in): fix ONLY a real box, never a title. Sample slide bg (median of edge points away from corners) and fill the corner with it only if: white_px > 1500 AND dark_px < 300 inside [0,0, 0.14W, 0.15H] (>244 all-channels = white; <120 all-channels = dark text). A left-aligned title gives ~200-400 "white" px (antialiasing) + lots of dark px — DO NOT fill, or you erase the first word ("2012:" → cut).
  • Also keep using the anti-box wording in any custom EXEC_PREFIX you copy into a deck's _common.py (the reference decks already do).
• Build a deck from full-bleed PNGs + ported notes (no LibreOffice): slide.shapes.add_picture(png,0,0,width=SW,height=SH) on blank 13.333×7.5" layout; port notes by position from a source pptx (deepcopy the BODY notes placeholder if target has none); export to PDF via PIL imgs[0].save(out,save_all=True,append_images=imgs[1:],resolution=150).
• Verify without poppler/LibreOffice: render PDF pages with PyMuPDF (fitz.open(p)[i].get_pixmap(dpi=80)); detect logo side via indigo-mask (b>110)&(b-r>40)&(b-g>30)&(r<140) count in top-left vs top-right band.
• Cyrillic: markitdown shows mojibake in the terminal (cp1251 console) — the file is fine UTF-8. Verify slide text visually (text is baked into images), not by terminal dump.

Modular production pattern for big decks (build.py + notes.py + assemble.py) ⭐

For real multi-slide decks (30–64 slides) WITH speaker notes, the maintainable scaffold is the modular deck pattern, not one giant config JSON. Готовых примеров-деков в паке нет (это локальные артефакты автора) — каркас ниже описывает паттерн целиком.

deck/
├── _common.py        # shared EXEC_PREFIX + make_client() + gen_slide() + run_generation()
├── assemble.py       # pptx from exec_raw/*.png + notes.py; supports --swap-speaker
└── slides/
    ├── build.py      # SLIDES = {n: '''english layout + "русский текст в кавычках"'''}
    ├── notes.py      # NOTES  = {n: '''разговорный спикерноут'''}
    └── exec_raw/      # slide_NN.png 1280×720 (gemini-3-pro-image-preview / nano-banana-pro)

• run_generation(SLIDES, RAW) skips PNGs already >100 KB → idempotent. Regenerate one slide: python slides/build.py --only=5,12. Edit a slide's text in build.py, delete its PNG, re-run --only=N.
• EXEC_PREFIX lives once in _common.py (cream #F1F3F5, marker icons indigo #3B5BDB / copper #C77B30; anti-corner-box, anti-Latin-lookalike Cyrillic «render И/С/Н as Cyrillic, NOT I/C/H», NO duplicates, render each element exactly once). Per-slide prompt holds ONLY content.
• python assemble.py slides [--swap-speaker] lays each slide_NN.png full-bleed on a blank 16:9 pptx and ports NOTES[n] into the notes slide (clone BODY notes placeholder via deepcopy when target has none). --swap-speaker overlays a ready speaker card onto slide_02.
• Insert / reorder slides WITHOUT regenerating everything (remap): keep the old build as build_v1.py; new build.py does from build_v1 import SLIDES as OLD, maps old→new index (e.g. shift keys ≥11 by +1) and adds NEW = {...}. Then shift PNGs to match — descending mv slide_30→31 … slide_11→12 — delete only changed numbers, and --only= the new/changed slides. Same trick for notes.py (programmatic remap script). Adds a mid-deck slide in ~30 s of gen instead of 11 min.
• PowerPoint lock: if the target .pptx is open, assemble raises PermissionError. Fall back to a new name (import assemble; assemble.OUT_NAMES['slides']='..._обновлено.pptx'; assemble.main('slides')) and tell the user to close the original.
• QA before delivery: montage exec_raw/*.png into 2–3 contact sheets (PIL, 4×4) and eyeball the whole deck.

Slide text + speaker notes: NO infobiz, human language ⭐ (YourFirstName's hard rule)

On-slide text AND speaker notes must read like a human, not an info-business coach. Strip on every pass:

• Empty value-less framing: «Инструмент мощный, но толк только если правильно», «честная рамка, чтобы не разочароваться».
• Hype adjectives: «прям клад», «огонь», «золото», «кайф», vague «в десять раз».
• Clichés: «лучшее время начать было…», «дисциплина быстрых циклов / Shipped > Perfect», «золотое правило».
• Garbled idioms: «не молотком по гвоздику» → plain («простую — лёгкой моделью, сложную — мощной»). Replace each with a concrete, useful statement or cut it. Bottom-marker lines should teach or instruct, not motivate.

Speaker-notes voice for YourFirstName's decks

Write notes.py in the speaker's real spoken voice — заведи собственный voice-profile файл (в пак он не входит: голос спикера — вещь личная) with concrete markers: metronome/filler openers for transitions, favourite connectives, peer-address forms, homely analogies, self-irony on glitches.

slide_manager.py (HTML Mode Core)

State machine for HTML-based projects. Manages slide_state.json.

slide_templates.py (HTML Templates)

13 named styles generating self-contained 1280x720 HTML slides.

slide_export.py (HTML Export)

Style Recommendation Guide

Prompt Engineering

Quality depends entirely on prompt quality. Key rules:

Structure

TITLE in large bold marker: "Заголовок слайда"
SUBTITLE: "Подзаголовок"

[Layout description with specific content]

At the bottom: "Key takeaway or footer text"

Content Elements

• Lists: "Numbered list: 1. First item 2. Second item"
• Diagrams: "Draw a flowchart: Box A → Box B → Box C"
• Charts: "Draw a bar chart: Year 1: 20M, Year 2: 100M"
• Tables: "Draw a table: Feature | Status | Priority"
• Icons: "Draw a shield icon", "a brain icon", "a clock icon"
• Highlights: "In a green box:", "In a red circle:", "Underlined in blue:"
• Comparisons: "LEFT side: ... RIGHT side: ..."
• Colors: black (headers), blue (subtitles), green (positive), red (critical), purple (accents)

Example Prompt

TITLE in large bold black marker: "8-слойная архитектура"
SUBTITLE in blue marker: "Снизу вверх — от души до интерфейса"

Draw a vertical stack of 8 colored horizontal layers:
Layer 8 (gray): "ADMIN PANEL — Control UI"
Layer 7 (cyan): "CHANNELS — Telegram, Slack"
Layer 6 (orange): "PROACTIVITY — Heartbeat, Cron"
Layer 5 (green): "TOOLS — Calendar, Email, Search"
Layer 4 (blue): "BRAIN — LLM Router"
Layer 3 (purple): "MEMORY — Qdrant, Hybrid Search"
Layer 2 (teal): "DATA — PostgreSQL, APIs"
Layer 1 (red): "SOUL — Constitution, Values"

Draw small icons next to each layer.
On the right: vertical arrow labeled "Уровень абстракции"

Mode Selection Guide

Unified Pipeline (Whiteboard Mode) — FOLLOW THIS

When user asks to create a presentation ("сгенерь презу", "create slides", "make a deck"), follow these steps:

Step 1: UNDERSTAND

Clarify topic, audience, goal, language. If user gave a clear topic — proceed without asking.

Step 2: STYLE

Run python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py recommend "<topic>" to get top style suggestions. Pick the best one or let user choose. Default: whiteboard.

Step 3: OUTLINE

Generate a JSON config with detailed visual prompts. Write it to a temp file:

{
  "title": "Presentation Title",
  "style": "whiteboard",
  "slides": [
    {"id": "slide_01", "prompt": "TITLE in large bold marker: \"Title\"\nSUBTITLE: \"Subtitle\"\nDraw relevant illustration below."},
    {"id": "slide_02", "prompt": "TITLE: \"Problem\"\nDraw diagram showing..."}
  ]
}

Rules for prompts:

• Each prompt must be 3-5 sentences minimum with SPECIFIC visual details
• Use Prompt Engineering section below for structure (TITLE, SUBTITLE, layout, colors)
• First slide = title, last slide = summary/CTA
• Include real data, numbers, examples where relevant
• 8-12 slides is optimal (no fewer than 6, no more than 15)

Step 4: REVIEW

Show the outline to user as a formatted list:

1. Title Slide — "Presentation Title"
2. Problem — diagram showing market gap
3. Solution — architecture overview
...

Wait for OK or edits. If user says "OK" / "давай" / "go" — proceed. If edits requested — update config and re-show.

Step 5: GENERATE

python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py generate <config.json> <output_dir> [style]

This auto-creates presentation.pptx + preview.html.

Step 6: CHECK

After generation, review results. If any slides failed or look bad (user reports), proceed to Step 7.

Step 7: REGENERATE (if needed)

python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py regenerate <config.json> <output_dir> <slide_id> [style]

This deletes the old image, regenerates, and rebuilds PPTX + HTML.

Step 8: SPEAKER NOTES (optional)

If user wants speaker notes ("добавь заметки докладчика", "add speaker notes"):

1. Generate notes as JSON: [{"index": 0, "notes": "Welcome everyone..."}, ...]
2. Save to notes.json
3. Run: python ~/.claude/skills/manus-slides/scripts/whiteboard_generator.py notes-pptx <image_dir> <output.pptx> <notes.json>

Step 9: DELIVER

Inform user:

• PPTX: <output_dir>/../presentation.pptx
• HTML preview: <output_dir>/../preview.html
• Individual slides: <output_dir>/slide_01.png, slide_02.png, ...

Research Mode

When the topic needs real data ("с актуальными данными", statistics-heavy topic, market analysis):

1. Use WebSearch to find current facts, statistics, trends
2. Collect key numbers and sources
3. Incorporate real data into slide prompts at Step 3 (OUTLINE)
4. Example: instead of "Draw a bar chart showing growth" → "Draw a bar chart: 2023: $150B, 2024: $185B, 2025: $220B (source: Gartner)"

URL Source Mode

When user provides a URL ("сделай презу по этой статье"):

1. Use WebFetch to read the URL content
2. Extract key points, data, structure
3. Use this content to generate the outline at Step 3
4. Reference source in the title slide

Change Style Flow

When user says "поменяй стиль на X" / "change style to X":

1. Update the "style" field in config.json
2. Delete all existing slide images in output_dir
3. Re-run generate with new style
4. All 24 AI styles are available (see AI Styles section below)

Legacy Pipeline (HTML Mode)

HTML Pipeline

1. Understand → 2. Outline → 3. Confirm
2. Init — slide_manager.py init
3. Generate — Write HTML per slide using templates
4. Export — slide_export.py html/pdf/pptx

Project Structure

Whiteboard Mode

my-presentation/
├── slides.json           # Config with prompts + style
├── generated/
│   ├── slide_01.png
│   ├── slide_02.png
│   └── ...
├── presentation.pptx
└── preview.html

HTML Mode

my-presentation/
├── slide_state.json
├── slides/
│   ├── slide_001.html
│   └── ...
├── presentation.html
└── presentation.pptx