Clean Slides

An opinionated CLI for AI agents to generate consulting-style PowerPoint slides from YAML. Not a general-purpose slide builder. It produces the kind of minimal, information-dense tables I preferred to write at McKinsey: no graphics, no effects, just structured thinking in a clear, visual format.

Here are two examples, neither touched by me:

(and see a fun one-shotted deck about the adorable Russian Desman here)

What it does

Your agent writes a YAML spec describing a table's content and structure. Clean Slides handles the layout — column widths, row heights, font sizing, dividers, bullet hierarchy — and produces a .pptx file that looks like it was built by hand.

title: The potential scale of tokenization adoption hinges on multiple factors.
subtitle: Forecasts of three potential adoption scenarios, 2030, nonexhaustive

table:
  rows: 6
  cols: 4
  has_col_header: true
  has_row_header: true
  col_headers: ["Slower adoption", "Base scenario", "Accelerated adoption"]
  col_header_color: accent1
  column_widths: [0, 1, 1, 1]  # custom weights (remove for html-4 style default)

  row_headers:
    - - { text: "Potential 2030 tokenized market cap" }
      - { text: "US$ trillion", color: tx1, bold: false }
    - "Regulation"
    - "Infrastructure and enablers"
    - "Market demand"
    - "Risk and reputation"

  row_overrides:
    0:
      align: ctr
      anchor: ctr
      bold: true
      color: accent1
      size: 20

  cells:
    - ["$0.8T", "$1.9T", "$3.8T"]
    - ["Regulatory challenges remain high", "Regional disparities exist, but regulatory clarity is increasing", "Permissive regulation"]
    - ["Inadequate infrastructure to support tokenization, lack of enablers (eg, institutional-grade stablecoin)", "Infrastructure reaches adequate maturity, required enablers emerge", "Infrastructure and enablers reach institutional-grade maturity, available at scale"]
    - ["No solution to cold start problem limits demand for tokenization", "Some assets tokenized at scale catalyzed by larger players", "Cold start problem is overcome by industry-wide collaboration"]
    - ["Systemic risk event occurs and limits or discourages adoption", "Limited security issues or systemic risk events", ""]

pptx generate tokenization.yaml -o tokenization.pptx

That's mostly it. The tool also handles:

• Inline formatting — **bold**, *italic*, [links](url) in any cell
• Bullet hierarchy — nested paragraphs with configurable indent levels
• Icon indicators — traffic-light circles for RAG status, severity, etc.
• Sidebars — split layouts (2/3, 3/4) with formatted text in the secondary area
• Slide inspection and editing — read shapes, edit text, merge decks

Charts (JSON) — alpha

Clean Slides also includes a chart generator for bar/stacked/waterfall charts driven by JSON. This is alpha — the interface works but the schema and CLI flags may change.

pptx charts spec.json output.pptx

See docs/CHARTS.md for the JSON schema and options.

What it doesn't do

This is a text table tool, not a presentation platform. It won't:

• Generate diagrams or images
• Apply transitions or animations
• Create freeform layouts or infographics
• Auto-generate content — you think the thoughts, the agent helps you structure them and handles formatting

The design is deliberately constrained. Tables are the primary medium because a table is the minimal way to express structured thinking. Even the default bullet style at level zero is no bullet — because bullets add visual clutter that doesn't serve the reader.

If you want a tool that produces polished decks with graphics and imagery, this isn't it. If you want something that takes structured text and produces clean, consistent, professional-looking table slides — fast — this is what it's for.

Template

• Quick start: works out-of-the-box with the bundled example template which is used for the screenshots above
• BYO: load your own template by dropping it in .clean-slides/template.pptx and creating the config

Quick Start

With Pi:

pi install npm:@tmustier/pi-clean-slides

Or manually:

# clone into your agent skills directory (e.g. ~/.pi/agent/skills/, ~/.claude/skills/)
git clone https://github.com/tmustier/clean-slides && cd clean-slides
pip install -e .
pptx init           # creates .clean-slides/ with a bundled example template

Write a YAML file (see example above), then:

pptx generate slide.yaml -o output.pptx

The template and config in .clean-slides/ are discovered automatically — no flags needed.

Using your own template

Any .pptx works as a template — a corporate deck, a downloaded theme, or something you design.

# Initialise with your template (auto-generates a starter config)
pptx init -t my-template.pptx

# Or do it manually
pptx init-config my-template.pptx -o config.yaml   # inspect + generate config
mkdir .clean-slides
cp my-template.pptx .clean-slides/template.pptx
cp config.yaml .clean-slides/config.yaml

Then edit .clean-slides/config.yaml to map your template's colours, fonts, and placeholder indices. See docs/TEMPLATE-CONFIG.md for the full reference, and examples/custom-template/ for a worked example.

Commands

Command	Purpose
pptx generate <yaml...>	Generate slides from YAML specs
pptx charts <json> <output>	Generate charts from JSON specs (alpha)
pptx validate <yaml...>	Check YAML against schema
pptx verify <yaml...>	Check sizing and overflow
pptx show <file> [slide] [shape]	Inspect slides and shapes
pptx layouts <file>	Show template layouts and content areas
pptx theme <file>	Show colour scheme
pptx edit <file> <slide> <shape> <text>	Edit shape text
pptx insert <deck> <source>	Merge slides from another file
pptx render <file> <slide>	Render to PNG
pptx init [-t template]	Bootstrap a .clean-slides/ project
pptx init-config <template>	Generate config from a template

Full YAML schema: docs/INPUT-SCHEMA.md

Development

pip install -e '.[dev]'
pre-commit install
python -m pytest -q
pyright clean_slides/

Why this exists

I spent six years writing slides in consulting. Most AI slide tools focus on imagery and visual polish — which has its place, but isn't what makes a consulting slide useful. A good slide is a unit of structured argumentation that happens to be visual. You figure out the narrative first, decompose it into pages, then fill each page with the component parts of your argument.

This tool takes that workflow seriously. You focus on the content and structure; it handles the formatting. When used with an AI agent, it produces first drafts that feel like they belong in the document — consistent style, proper hierarchy, no "generated by AI" aesthetic. The constraint is the point.

More on the design philosophy: mustier.ai/projects/clean-slides

Clean Slides works with any agent, but I use pi.dev with my Agent Teams extension that lets the agent create clones of itself with the same context and manage them — so each clone worked on one page with full context, and the original agent focused on output quality and consistency, similar to how a consulting team would split work after a group problem-solving meeting. You can optionally do this with:

npm install -g @mariozechner/pi-coding-agent
pi install npm:@tmustier/pi-agent-teams