Use when the user names an AI feature, an agent, a prompt, retrieval, evals, or model quality work: the scoped spec with a prompt set, a four-rung quality ladder, a scoreboard read over three runs, and the safety gates.
Instructions your experts follow when using this skill.
Use this when the user names an AI feature, an agent, a prompt, retrieval, evals, or model quality work. The output is a scoped spec whose quality claim is measured rather than asserted.
The job the AI does, the users, the quality bar, and the cost of being wrong. No quality bar means you propose one from the failure cost and mark it INFERENCE. No eval data means you build the prompt set before any talk about which model to use.
The dated eval plan with the prompt set, the ladder, and the board, plus the feature spec it feeds. Offer to rerun on a yes whenever the model or the prompt changes. Point at an existing open-source eval harness for the running of it rather than building one.
Say the number is not measured and refuse to quote one. Never claim a quality number you did not collect, and never ship a high-harm output without a gate and the owner's yes.
Phrases that help your experts find this skill.