The thinking behind agent-slides.
Design decisions, architecture patterns, and lessons learned building an agent skill for PowerPoint generation.
Why 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 → 02JSON payloads and runtime schemas
Why CLIs for agents should accept structured data instead of flags, and why the API contract should come from the tool itself.
Read → 03Context-window discipline
Why output volume is a hard design constraint, and why deterministic output is non-negotiable for automation.
Read → 04Agent output as untrusted input
Why agent-driven CLIs should validate like a public API, and why design taste must become machine-readable contracts.
Read → 05Skills encode expertise, not instructions
Why --help isn't enough. Skill files give agents the workflow knowledge that documentation can't convey.
Read → 06Why seven skills, not one
Why one monolithic skill became seven focused ones, and why conditional reference loading keeps context lean.
Read → 07File contracts and state-machine orchestration
Why skills communicate through files on disk, and why the orchestrator converges through retry loops instead of a linear pipeline.
Read → 08Progressive disclosure for agents
Give the agent a map, not a manual. Each layer reveals only what the current step needs, so context stays lean and decisions stay grounded.
Read →