slide-skills 0.2.2

AI slide generation skills — fill .pptx/SVG templates with AI text, images, themes, and animations

Slide Generator Lab

AI slide generation in Python. Take a template (a Canva/PowerPoint .pptx, or a hand-designed SVG collection), and generate a finished deck whose text, colors, images, and animations are produced by AI to fit a topic — while the layout stays exactly as designed.

Built as composable skills so the whole thing can be wired into a FastAPI app. Published on PyPI as slide-skills.

---

Install

As a library (use in any project):

pip install slide-skills                 # latest release from PyPI
# or, the bleeding edge straight from source:
pip install "git+https://github.com/phatgg221/Slide-generator-lab.git"

For development (edit the code, changes apply instantly):

git clone https://github.com/phatgg221/Slide-generator-lab.git
cd Slide-generator-lab
python3 -m venv .venv && source .venv/bin/activate
pip install -e .                         # editable install

Optional extras:

pip install "slide-skills[svg-convert]"  # PyMuPDF, for .pptx -> SVG conversion
pip install "slide-skills[all]"          # everything optional

Requirements

What	Needed for	Notes
OPENAI_API_KEY	any AI step (text, images, planning)	put in env or a .env file
Python ≥ 3.9	everything
resvg-py (auto-installed)	rendering SVG/web decks	bundled, no system deps
LibreOffice	svg_template_maker only (.pptx → SVG)	brew install --cask libreoffice; optional

Environment variables (all optional):

OPENAI_API_KEY=sk-...            # required for AI calls
OPENAI_TEXT_MODEL=gpt-4o         # default
OPENAI_IMAGE_MODEL=gpt-image-1   # skip image-model auto-detection
SLIDE_TEMPLATES_DIR=/path/to/svg/templates   # where your SVG collections/categories live
SLIDE_LIBRARY_DIR=/path/to/pptx/templates    # where your .pptx templates live

The two *_DIR vars are the key to using this as an installed package: set them once and the library finds your templates wherever you keep them — your designs live outside the code package.

---

Two paths

The project supports two delivery targets that share most of the same skills:

	PPTX path	Web path (current focus)
Output	Downloadable PowerPoint file	Self-contained animated HTML for your website
Template source	Canva/PowerPoint .pptx	Hand-designed SVG collections (Figma/Inkscape)
Animation	PowerPoint transitions + entrance effects	Native SVG/CSS animation (Canva-like)
Main CLI	examples/build_course_deck.py	examples/web_deck.py

The web path is preferred because SVG keeps text editable, colors remappable, and animations playable in the browser — and it needs no desktop renderer.

---

Use as a Python library

Once installed, import the skills from anywhere:

from slide_skills import generate_web_deck

# Topic -> animated HTML deck (writes the file, returns a summary dict)
result = generate_web_deck(
    collection="starter",                 # folder under SLIDE_TEMPLATES_DIR
    brief="Khóa học nhập môn Machine Learning",
    output_path="out/deck.html",
    palette="teal",                       # "auto" | preset name | (primary, secondary, accent) | None
    language="Vietnamese",
    animation="rise",                     # rise | fade | scale | none
)
print(result["output_path"], result["slides"])
print(result["usage"]["report"])          # tokens + estimated USD for this call

Every generate_* call returns a usage block with the total cost of that run:

result["usage"]   # {input_tokens, output_tokens, total_tokens, requests,
                  #  estimated_cost_usd, report}

Category library + variant-selecting agent (your plan names categories; the agent picks the best design variant per slide):

from slide_skills import generate_deck_from_plan, scan_template_library

lib = scan_template_library("templates")     # discover categories + variants
print(lib.category_map())                     # registry for a UI / planner

plan = {"title": "ML 101", "slides": [
    {"category": "Title Slide", "topic": "Intro to Machine Learning"},
    {"category": "KPI & Big Numbers", "talking_points": ["78%", "3x", "12M"]},
    {"category": "Conclusion & Summary", "talking_points": ["Recap", "Next steps"]},
]}
generate_deck_from_plan(plan, "templates", "out/deck.html", palette="auto", language="Vietnamese")

Add a collection at runtime (e.g. a user uploads a Figma export):

from slide_skills import import_collection
import_collection("/path/to/figma_export_folder", "my_style")   # copies + validates

Track cost of any run:

from slide_skills import usage_tracker
before = usage_tracker.snapshot()
# ... generate ...
print((usage_tracker.snapshot() - before).report())

In a FastAPI app, wrap blocking calls in a thread:

import asyncio
from fastapi import FastAPI
from fastapi.responses import FileResponse
from slide_skills import generate_web_deck

app = FastAPI()

@app.post("/generate")
async def generate(topic: str, collection: str = "starter"):
    out = "out/deck.html"
    await asyncio.to_thread(generate_web_deck, collection, topic, out)
    return FileResponse(out, media_type="text/html")

(Set SLIDE_TEMPLATES_DIR so the app finds your collections.)

---

Quick start (CLI) — Web deck from an SVG collection

# 1. See available collections
.venv/bin/python examples/web_deck.py list

# 2. Validate a collection — what placeholders did it find? (free, offline)
.venv/bin/python examples/web_deck.py check starter

# 3. Visual preview with stub text (free, offline)
.venv/bin/python examples/web_deck.py demo starter
open out/starter_demo.html

# 4. Generate a real deck from a topic (~$0.02)
.venv/bin/python examples/web_deck.py generate starter \
    "Khóa học nhập môn Machine Learning" -o out/ml_deck.html --language Vietnamese
open out/ml_deck.html

In the browser deck: →/← to navigate, f for fullscreen, elements animate in as each slide appears.

Options: --palette teal (force a theme), --animation rise|fade|scale|none, --pptx (also export a PowerPoint copy).

---

Designing your own SVG collections

You design collections once in Figma/Inkscape; every generated deck reuses them. A collection is a folder of slide-type SVGs sharing one visual style:

svg_templates/<collection>/
  collection.json     # optional: description, palette, fonts, tags
  title.svg           # filename = slide type the planner picks from
  statistic.svg
  comparison.svg ...

Rules (see svg_templates/README.md for the full guide):

• Export SVG with "Outline Text" UNCHECKED — text must stay live <text>.
• Placeholders are the text content: {{title}}, {{quote|120}} (120-char budget), {{body.1}} / {{body.2}} (multi-line).
• One uniform style per placeholder (don't bold half a word — it splits the text).
• Name files by function (title, statistic, quote…); use common fonts.

Then web_deck.py check / demo your folder before spending on generation.

---

Category library + variant-selecting agent

For richer decks, organize templates into categories, each holding several variant designs for the same layout function. The user's plan names a category per slide; an agent picks the best-fitting variant for the data.

templates/
  TITLE_SLIDE/
    category.json        # optional: descriptions guiding variant choice
    centered.svg         # variants — multiple designs, same purpose
    left_aligned.svg
  KPI_BIG_NUMBERS/
    category.json
    three_stats.svg
    list_style.svg

category.json (optional but recommended — it's the "map" the agent reads):

{
  "description": "Highlight key metrics with large, scannable numbers.",
  "variants": {
    "three_stats": "Three stat callouts — use for 2-3 key numbers.",
    "list_style":  "A vertical list — use for 4+ numbers or rankings."
  }
}

How a variant gets chosen, per slide:

1. schema-fit shortlist (code) — keep variants whose slot count fits the data
2. AI tiebreak (GPT-4o) — read each finalist's description + slots and the slide content, pick the best, and write its text

Adding designs is pure data:

• New variant → drop a .svg in the category folder (instantly usable; add a category.json line so the agent knows when to choose it).
• New category → make a new folder; it appears in category_map() automatically.

Category names match plan labels ignoring case/spaces/&/-/_ ("KPI & Big Numbers" → KPI_BIG_NUMBERS), but not plurals — name folders to match your plan labels.

---

Quick start — PowerPoint deck from a .pptx template

# Ingest any .pptx into the reusable template library (cleans junk, classifies slides)
.venv/bin/python examples/prepare_template.py "~/Downloads/My Design.pptx" my_template

# See what's editable (free, offline)
.venv/bin/python examples/test_template.py my_template

# Full pipeline: research -> plan -> write -> images -> theme -> animate
.venv/bin/python examples/build_course_deck.py library/my_template.pptx \
    "your topic" --transition fade --animate fade -o out/deck.pptx

Cost controls: --no-research, --no-images, --svg-images (cheap vector illustrations instead of AI photos).

---

Skills reference (slide_skills/)

Foundation

• config.py — OpenAI client + model config from .env
• usage.py — token & cost tracking across all AI calls (usage_tracker)
• template_parser.py — parse a .pptx into a fill-spec; classify text roles, char budgets; skip tip-bubbles & navigation buttons

Research → Plan → Write

• research.py — extract_keywords, web_research
• planner.py — plan_deck: AI picks slide count, types, order, theme
• content_generator.py — generate_content: AI writes budget-aware text
• agent.py — SlideGeneratorAgent: fill one template from a brief
• pipeline.py — CourseDeckPipeline: the full chained pipeline

Images

• image_generator.py — generate_image: AI photos, auto-detects account's model
• svg_image_generator.py — generate_svg_image: GPT-4o vector art (~5× cheaper)

Filling & assembly

• slide_filler.py — write text keeping formatting, auto-shrink overflow, swap images
• assembler.py — build a deck by picking/reordering/repeating library slides

Templating

• template_maker.py — prepare_template: ingest + clean + AI-classify a .pptx
• merge_template.py — {{placeholder}} form + schema; AI fills; render
• svg_template_maker.py — .pptx → folder of live-text SVGs (needs LibreOffice/PowerPoint)

Theme & motion

• theme.py — contrast-safe recoloring, 8 presets, propose_palette
• transitions.py — PowerPoint slide transitions
• animations.py — PowerPoint element entrance animations

Web decks

• svg_collections.py — scan collections, fill placeholders, retheme, import_collection, generate_web_deck
• svg_categories.py — category library + variant-selecting agent: scan_template_library, select_and_fill_slide, generate_deck_from_plan
• html_deck.py — build a self-contained animated HTML presentation
• svg_slide_renderer.py — filled SVGs → PNG → .pptx export

---

Command-line tools (examples/)

Command	Purpose
web_deck.py	SVG collections → animated web deck (list/check/demo/generate)
build_course_deck.py	Full pipeline → .pptx
generate_deck.py	Fill one template from a brief
prepare_template.py	Ingest a .pptx into the template library
test_template.py	Dry-run marker fill — see what's editable (free)
recolor_deck.py	Re-theme an existing deck's colors
merge_deck.py	{{placeholder}} workflow (make/render/generate)

---

What can be customized per deck

• Words — every {{placeholder}} is AI-written, any language
• Colors — preset, AI-picked, or custom; always contrast-safe
• Animation — rise / fade / scale / none (web) or PowerPoint effects (pptx)
• Slides — the planner chooses which template types to use, and their order

Fixed by design: your layout, and (for now) fonts.

---

Tests (offline, no API key)

.venv/bin/python tests/test_offline_pipeline.py   # parse / fill / image swap
.venv/bin/python tests/test_assembler.py          # library assembly

---

Known limits

• Canva exports lose animation and live text. .pptx from Canva is static; Canva SVG outlines all text. Design real SVG templates in Figma/Inkscape.
• svg_template_maker.py needs a renderer. Install LibreOffice (brew install --cask libreoffice) for headless, server-ready conversion; the desktop-PowerPoint fallback is fragile.
• Charts aren't data-driven yet. Chart-style slides render as designed art, not recomputed from numbers.
• FastAPI app is a later milestone; all skills are import-ready for it.

---

Releasing a new version (maintainers)

# 1. bump version in BOTH pyproject.toml and slide_skills/__init__.py
# 2. rebuild fresh and publish
rm -rf dist && python -m build
twine upload dist/*          # username: __token__   password: your pypi-... token
# 3. tag it
git tag v0.2.2 && git push origin main --tags

PyPI versions are permanent — never reuse a number; bump to the next one.