The Elicitation Stack 🧱

One-liner: A capability system for engineering teams to get genuinely good at using AI — beyond “we adopted some tools” — built around a documented failure taxonomy, an observable competency ladder, and a monthly measurement loop.

Origin: Grew out of an internal AI-competency system built for my own engineering org, then generalized into something any team can adopt. Grounded in system-card research on how models actually fail in real usage, not vendor benchmarks.

Core idea:

Output quality = Model × Context × Tools × Human judgment × Verification loop

Multiplicative — a zero anywhere zeroes the result. Most orgs spend all their attention on the one layer (the model) they don’t control.

What’s in it:

• A failure taxonomy — six named, documented patterns (skipped verification, reported-as-done, fabrication, and more) that become a team’s shared vocabulary in code review
• A competency ladder, L0 (chat user) → L4 (systems owner), assessed by artifact review and adversarial defence — not self-reported
• A measurement loop — monthly transcript audits, one shipped countermeasure per audit, throughput and review burden always reported as a pair
• A build order — data boundary → task inventory → context → tools → verification → automation, in that order, because most teams build it backwards

Status: Actively running — I use this to run my own team’s AI-capability development. Developed as a full networked knowledge base alongside this public summary.

Stack: Markdown + Obsidian for the working system; this page is the public-facing overview.