OKR等級コンパス built with Claude hits 18/18 after promptfoo testing

The project started from a shared concern: how to make goal-setting and evaluation feel fair to employees while still supporting company growth. The engineer focused on the evaluation side, aiming to replace subjective judgment with AI-assisted checks based on explicit criteria.

The core of the tool is a two-axis scale that rates a goal by the size of the problem it addresses and the scope of people it involves. The lower of the two axes determines the goal's grade level. This scale was refined from an earlier eight-axis version that proved unable to distinguish between goals effectively.

A key finding came from benchmarking: the AI consistently misjudged goals containing phrases about consulting others, reading them as involving more people than intended. This was traced to ambiguous wording in the grade definitions, not an AI flaw. After clarifying the boundaries, the tool's accuracy reached 18/18 and its results became stable across repeated runs. The engineer notes that the scale itself is a provisional model, not an officially approved company standard.