Handshake

A brief mutual exchange that establishes the terms of subsequent collaboration. Not a friendship, not an interview — a handshake.

This skill exists because agents that don't know who they're talking to give generic answers. It exists as a separate skill — rather than relying on passive memory accrual — because deliberate calibration is structurally different from incidental capture: the user gets to see what's already known, correct it, and consent to what's added. That's the entire job.

The three-part test (front and center)

Per the agent-skills repo convention, every new skill is checked against three axes before it earns a slot. Here's where handshake lands honestly:

Axis	Verdict	Notes
Unique structure	Partial	Show-then-ask is a different control flow from passive memory accrual, but the underlying memory store is shared with the auto-memory system and `productivity:memory-management`.
Unique deliverable	No	The output is standard `user`-type and `project`-type memory entries — same format other surfaces produce. The skill wraps the existing memory contract; it does not invent a parallel one.
Unique elicitation	Yes	Privacy-conscious, consent-gated, show-what-I-know-first, ≤4 pill + 1 free-text core, optional ≤6 project overlay, hard never-ask list. This is the load-bearing reason the skill exists.

The skill ships because the elicitation pattern is genuinely worth ritualizing — but it stays narrow on the other two axes by deliberate design. If you find yourself extending it to invent new memory types or invent new control flows, stop and ask whether you're really still inside handshake.