Tools
The Median Dial
An LLM's unconstrained output is the statistical median of everything it has read — which is exactly why it reads like nobody in particular. This dial runs the same brief through three fixed states, for three media: the median output an AI produces with no constraints, a cleaned draft with the buzzwords removed, and a director's cut written under a handful of specific constraints. Drag it and watch two gauges track the distance from the median in real time.
Brief: write the opening screen of a new-manager training module on giving feedback.
Welcome to this transformative learning journey! In this dynamic module, you'll leverage cutting-edge, best-in-class frameworks to unlock your full potential as a people leader. We'll seamlessly empower you to foster a culture of continuous feedback, driving impactful, actionable outcomes across your team. By the end of this holistic experience, you'll be equipped with robust, synergistic tools to elevate every conversation and champion a growth mindset organization-wide.
Directorial constraints
- No adjective may appear without a specific number, name, or action attached to it.
- State the outcome the learner will be able to do, not how they will feel.
- Every sentence must survive being read aloud to the person it's about.
The Corporate Glaze Index and Lexical Surprise Score are computed live from the text on screen, in your browser — no model call, no hidden ranking. Both are built from the same handful of measurements: buzzword and corporate-trope density, Latinate abstraction density, adverb density, sentence-length uniformity, and vocabulary commonness against a curated common-word list. Glaze weights those toward generic; Surprise weights vocabulary rarity, sentence-length burstiness, and monosyllable ratio toward specific, and subtracts the buzzword penalty.
The three states are hand-authored, not model-generated on the fly — on purpose. A live “try your own brief” version could embarrass the point by occasionally producing something good at 0%. These three don't, because they're fixed: the median output is deliberately the laziest true version of what an unconstrained model tends to produce for that brief, and the director's cut is what the same brief looks like once you write specific constraints into it instead of asking for “better.”
The move this tool is arguing for works outside training content, too: pair it with the prompt library for prompts that build constraints in up front, instead of generating the median first and editing it out later.