Skills is a repository of reusable AI agent skills focused on the craft of building technical presentations. An orchestrator skill detects what phase a project is in and guides through the next step; phase skills can also be invoked individually. The project is an experiment in making AI-assisted workflows portable, composable, and predictable across sessions.
Each skill is a Markdown document that an AI agent reads and follows. The orchestrator (build-presentation) inspects the project directory, detects what state the work is in — no discovery yet? agenda not approved? slides ready? — and invokes the appropriate phase skill. Phase skills can also be run directly when the user knows exactly where they are. State lives in plain files: DISCOVERY.json captures requirements, AGENDA.md holds the structure, and shared modules enforce a common schema across skills.
This project came from frustration. Every time I built a technical presentation with AI help, I was retracing the same steps — explaining context, defending choices, recovering from drift. Skills is an attempt to encode that process so the agent has a script to follow, not a tabula rasa each time.
The orchestrator pattern was the breakthrough. Letting one skill detect state and route to the right phase skill means the user can drop in at any point and the agent knows what to do next. It also makes each phase skill simpler — they only have to do one thing well.
Writing skills is more like writing a runbook than writing code. The hardest part is being explicit about decisions that humans make implicitly: what counts as "done" for discovery? When is an agenda ready to generate from? Each ambiguity in the prose surfaces as agent drift in practice.