The index

A program that uses an LLM to take actions against a goal — not just generate text.

Diagnosing whether a stalled project needs technical work, ownership work, or narrative work — and refusing to do all three at once.

Reframing evaluation as a binding agreement between engineering, business, and compliance — rather than a test suite nobody owns.

Using two or three input/output examples as a task specification, in place of rule-based instructions.

Explicit constraints on an agent's behavior; the antidote to 'it should figure it out autonomously'.

Shipping a v1 with hardcoded data beats designing a perfect v∞ that never ships.

A stateful workflow framework for LLM applications built around cyclic graphs.

The state where an AI initiative has technically succeeded in a demo but has no clear path to production use.

A one-page document naming the specific failure modes, acceptable rates, measurement windows, and automatic actions for an AI system.

System prompts that specify what the assistant is forbidden to do, with concrete examples. Does more work than telling it what to do.

Treating the model's response as a typed API payload rather than free-form text to be heuristically parsed.

Start any eval design by listing the three failure modes you would personally be fired for. Those are the contract's load-bearing clauses.