The premise
Creativity is not a property like temperature, mass, or speed — it cannot be measured absolutely. It is a situated judgement. A work can be highly creative compared with school essays and unremarkable compared with avant-garde literature; creatively strong for a corporate brand and weak if the author knowingly copied a specific source.
So this tool never claims to tell you "how creative" something is in an absolute sense. Its claim is narrower and more honest:
This tool estimates how creative a work appears under a declared frame of comparison.
That is why it always asks four questions before judging: Creative relative to what? Creative for whom? Creative under what constraints? Creative as an artefact, or creative as a human act?
The research behind it
The design draws on several strands of creativity research:
The standard definition. Researchers broadly agree creativity involves producing something original and effective / useful / appropriate (the APA defines it similarly). But "original to whom" and "useful for what" are context-dependent — hence the frame.
The Consensual Assessment Technique (Teresa Amabile). The "gold standard" in creativity research: suitable judges rate works independently, accepting that creativity judgement is partly subjective and socially situated. The Lens plays the role of one such judge — explicitly instructed to be honest rather than flattering, and to show its reasoning.
Computational creativity metrics. Design-creativity research (e.g. Grace et al.) evaluates work along dimensions like novelty, value, and surprise rather than a single number. The Lens extends this to seven dimensions.
Originality-at-scale methods. Tools like Open Creativity Scoring and MuseRAG estimate originality by how far a response sits from common or prior responses. The Lens approximates this with its related-works map: surfacing what the work plausibly resembles or descends from.
Big-C / little-c distinctions. Everyday creativity, professional creativity, and historically significant creativity are different standards. Declaring the domain and author experience lets the Lens judge against the right one.
The seven dimensions
Each is scored 1–10 with a written rationale. They are deliberately separate — a work can be highly novel and barely coherent, or perfectly coherent and entirely derivative.
Novelty — how different is this from related works in its domain?
Value / effectiveness — does it work for its stated purpose and audience?
Surprise — does it violate expectations in a meaningful way?
Coherence — is it internally consistent, resolved, well-formed?
Transformation — does it meaningfully rework prior material, or merely echo it?
Constraint handling — did the creator do something interesting within limits?
Domain situatedness — is it new within the relevant field, not just new in general?
Scoring is calibrated to be honest: most work is moderately creative; 9–10 is reserved for work that would surprise experts in the domain.
Artefact vs. author — the key distinction
The most important idea in the Lens is the split between two different things people mean by "creative":
The creativity of the artefact — how novel, valuable, surprising, coherent and distinctive the work appears to a viewer.
The creativity of the author's act — how much transformation, judgement, risk, insight and synthesis the author actually contributed, given what they knew.
A work can look creative to a viewer who has never seen its influences. But if the author had the source material in front of them and lightly altered it, the authorial act is weak. Conversely, a work may look ordinary to outsiders but represent genuine independent invention.
That's why the Lens shows you a related-works map and asks which influences the author was aware of:
• Resemblance to a known influence weakens the authorial act — unless the work meaningfully transforms it.
• Resemblance to unknown works does not weaken the authorial act — that is independent invention.
The pipeline
1. Submission. You provide the work plus the frame: domain, audience, goal, constraints, author experience, and whether AI was used. If the frame is left empty, the model infers the most plausible one and names it — so the comparison is always explicit.
2. Analysis call. The work and frame are sent to a Claude model (your choice; Opus 4.8 by default) with a fixed system prompt encoding the principles above. The response is constrained to a strict JSON schema: seven scored dimensions with rationales, 3–7 related works each tagged as likely influence / close neighbour / common trope / genre convention / structural template, and a narrative reading in plain prose.
3. Influence declaration. You mark which related works the author was aware of. This is the human-in-the-loop step no fully automatic tool can do.
4. Verdict call. A second model call receives the original profile plus your declarations and produces the artefact-vs-author split with a final narrative. Each recomputation is kept — nothing is overwritten.
5. Record. Every analysis records the date, model used, app version, and token usage, and can be exported as a Markdown report or raw JSON.
Honest limitations
The related-works map is the model's judgement, not a database lookup. The model names what the work plausibly resembles from its training knowledge. It can miss obscure prior art and occasionally over-claim resemblance. A future version may use real semantic retrieval over an indexed corpus.
One judge, not a panel. The Consensual Assessment Technique uses multiple independent judges; the Lens currently uses one model. Running the same work through different models (selectable in settings) is a rough approximation of a panel.
Scores are ordinal, not interval. A 8 is not "twice as creative" as a 4. Read the rationales and narratives; the numbers are anchors, not measurements.
Authorial creativity is an estimate. It relies on your honest declaration of known influences and the model's inference about transformation. It cannot see inside the author's process.
Privacy & architecture
The Lens is a static page. Your Anthropic API key is stored only in your browser's localStorage and requests go directly from your browser to the Anthropic API — no intermediary server sees your key or your work. Analyses exist only in your browser session until you export them.