Generative AI has two dimensions.
You're only measuring one.
The Dual Condition framework decomposes prompt quality into Template Specificity (T1) and Generative Headroom (T2) — two independent dimensions whose interaction predicts output quality.
Specificity ignites the engine. Unknowability fuels it. You need both.
How precisely the prompt activates deep patterns. Domain vocabulary, structural constraints, persona, intellectual tradition. Vague inputs activate nothing.
How much of the output space the prompt leaves open. If you fully constrain the output, the model becomes a lookup engine, not a generator.
Quality is the product, not the sum. A prompt that maxes one but kills the other fails — accurately, mechanically, or vaguely. The sweet spot is a hyper-specific key opening into an infinite room.
One scoring engine. Open source. Free.
Paste any prompt. Get T1, T2, interaction score, zone classification, named signals, and an actionable recommendation. Runs entirely in your browser.
Open the analyzer →The four-channel heuristic scoring engine as a zero-dependency JavaScript package. MIT licensed. ~20 KB. 9 calibration tests.
View on npm →Agent SDK with first-party LangChain adapter. Score, gate, and monitor instructions inside multi-agent workflows. Audit mode by default.
View on npm →Predict prompt quality before you spend the compute.
Integrate T1/T2 scoring into your CI/CD pipeline. Evaluate every system prompt change before it reaches production. The heuristic scorer is client-side and free; the enterprise API with full Claude-powered analysis is in development.
A formal model with twenty falsifiable predictions.
Built on predictive processing, anchored autopoiesis, and structural keyframing. Connects human cognition and artificial generation as instances of a single generative architecture in different substrates.
Read the paper on Zenodo →Get notified when the API opens.
The full Claude-powered T1/T2 analysis API is in development for enterprise teams running production LLM workloads. CI/CD integration, audit-mode quality gates, anchor decay monitoring. Drop your email and you'll get one note when the beta opens.
One email at launch. No marketing. Unsubscribe with one click.