Agent hosts
MCP / skills consumers
Cursor, Claude, VS Code, custom agents: taste_search → taste_get → generate → taste_score. Same packages as Impeccable-class skills, but many standards, tagged and citable.
Applications
Practice · Institutions · Leverage
Where agents load judgment
Open packages earn their keep when agents and humans use them in real loops— gen-UI, writing, research, education—not only as static moodboards.
Situations
Product orgs
House standard against generative default
Teams drowning in AI drafts need a shared reject function. Publish a house package; score PRs for UI copy and empty states; fail CI on hard constraints. Transfer the same package into marketing without a second religion.
Editorial
Desk style as executable policy
Magazines and research blogs already have house styles—half unspoken. Profiles make onboarding non-telepathic. Score pitches and drafts; keep anti-exemplars from the reject pile as living curriculum for new writers.
Research & labs
Figures, abstracts, claims
Scientific taste is real: figure discipline, claim hygiene, statistical manners. A lab profile can score internal drafts before peer review— not to automate science, but to institutionalize “we don’t do that here.”
Education
Apprenticeship without proximity
Teachers externalize standards students can practice against between critiques. The point is not automated grading of art; it is high-frequency formative feedback aligned to a named school of thought.
Personal practice
A constitution for one’s own output
Individuals version a private profile: what they are training themselves to prefer. Score weekly work. Diff the constitution yearly. Taste becomes a practice log, not a personality myth.
OSS curators
Publish a named standard
Designers, writers, labs: ship a package with tags, license, and anti-exemplars. Others fork and cite. Later commercial registries can host private packs without changing the open format.
Researchers
Operationalize a paper
Turn an aesthetics or preference-learning result into a package + golden set. Agents load the operationalization; evals keep it honest.
Minimal workflows
WEEKLY
Reject pile → anti-exemplars
Every Friday, add two failures that almost shipped. Re-encode. The border sharpens where work actually happens.
REVIEW
Human rules, machine pre-sorts
Score a batch; humans only deep-read the ambiguous band (e.g. 0.4–0.7). Preserve attention for the cases that teach.
LAUNCH
Transfer before channel sprawl
Before opening a new surface (docs, sales, onboarding), transfer the profile and write three sample artifacts under the guidance.
HIRE
Portfolio against the standard
Candidates remain people. Optionally, score work samples against the house profile as one structured input—not a gate dressed as destiny.
Anti-patterns
Score as HR weapon
Using opaque taste scores for performance punishment destroys trust and trains compliance theater. Publish reasons; allow appeal; separate coaching from firing tools.
One global profile
Organizations need dialects: research vs. growth, internal vs. external. Composition and lineage beat a single monoculture profile.
Never updating anti-exemplars
Last year’s slop is not this year’s. Frozen anti-exemplars become folklore while new failures waltz through.
Replacing critique
If the score ends the conversation, the system has failed. Scores should start better arguments, not end them.
Adjacent systems
Taste pairs naturally with exploration tools—an adjacent possible map can propose moves; a taste profile can filter which moves feel like you or this institution. Likewise: design systems (tokens), eval harnesses (regression), and research OS tools (claims and evidence) supply complementary structure. Taste is the prior over form; those systems handle state, truth, and change.