Open-source agent skill
Your AI agent for
client-ready PowerPoint decks.
An agent skill that generates professional, brand-compliant presentations — powered by a production-grade CLI with validation, QA, and deterministic output.
Brand-compliant decks
Feed it your corporate template. The skill extracts layout rules, color zones, and font contracts — then enforces them on every slide it builds.
Validated before delivery
Every slide operation is schema-validated. A QA engine checks font sizes, contrast, overlaps, and bounds before the deck is finalized.
Deterministic output
Same input, same deck. No random IDs, no timestamp drift. Reproducible results you can trust in client-facing work.
From brief to shipped deck.
Extract your template
Run /slides-extract on any .pptx template. The skill analyzes layouts, colors, fonts, and spatial zones into a machine-readable contract.
Build the deck
Describe what you need. /slides-build plans the narrative, selects layouts, and compiles slide operations into your presentation.
Audit & polish
The skill automatically audits formatting, reviews storytelling quality, and adds speaker notes — before the deck reaches anyone.
Or run /slides-full to orchestrate all stages as a single pipeline with retry loops.
Built for agent DX, not human DX.
Human-first CLIs optimize for discoverability. Agent-first CLIs optimize for predictability and defense-in-depth. The skill is powered by a purpose-built CLI designed for the latter.
The agent is not a trusted operator
Agents hallucinate. The CLI treats every input as untrusted: Pydantic validation, control-character rejection, path-traversal checks, and transactional execution that rolls back on failure.
Schema introspection replaces documentation
Static docs stuffed into context are expensive and go stale. The CLI exposes every operation, parameter, and response type as queryable JSON at runtime. The agent self-serves.
Context-window discipline
Agents pay per token and lose reasoning capacity with irrelevant output. Field masks, compact mode, and structured JSON keep every response minimal and machine-readable.
Skills encode what --help can't
Seven skill files encode workflow knowledge agents can't intuit: when to dry-run, how to chain stages, how to recover from errors, and what design invariants to enforce.
Install, extract, build.
# Install skills for your agent (Claude Code, Cursor, Gemini CLI, Codex)
npx skills add https://github.com/mpuig/agent-slides
# Extract your corporate template
/slides-extract corporate.pptx
# Build a full deck from a brief
/slides-full "10-slide strategy deck on Q3 growth"The engineering behind the skill.
Getting Started
Install, extract a template, and build your first deck. From prompt to polished presentation in minutes.
Read → 02Why agents need a different CLI
Human DX and Agent DX are different problems. Why the agent needs a dedicated execution layer, not raw library access.
Read → 03All design notes
Design notes on JSON payloads, schema introspection, context discipline, progressive disclosure, and state-machine orchestration.
Browse →