Scoville Design Anti AI Slop

Scoville Design Anti-AI-Slop

Make visual decisions from the subject and task. Composition foundations live here; load specialists only for an open decision or a needed method.

Mode, authority and ownership

Design owns concept, art direction, visual hierarchy, typography, composition, spacing, responsive intent, corporate-design/visual-identity definition and critique. Preserve incumbent systems and protected content. Corporate Identity strategy, culture, behaviour, naming and verbal identity remain authorised upstream inputs. Never invent facts, approval, rights, strategy or deliverables.

With optional active-applicable UI, Design defines and judges; that partner implements framework/component mechanics, states, semantics, accessibility, responsiveness and runtime proof. Use Coordination for an actual partner exchange (including Scribe for wording fidelity). Never require or simulate another Skill; without a partner, handle authorised work directly within the same ownership floors.

Finding another Skill in this family does not make it installed, active, applicable, or required. If that Skill is absent or inactive, ignore it. Do not require, install, simulate, or reimplement it. If it is active and applicable, let it handle only its stated concern while this Skill continues its own authorized work. An opt-out applies only to the Skill the user excluded, not to independently authorized work.

Family owners, in suite order:

• scoville-brainstorm: deliberate divergence before selection.
• scoville-research: source-backed research and synthesis.
• scoville-code-anti-ai-slop: engineering scope, implementation, risk, and validation.
• scoville-design-anti-ai-slop: visual definition, art direction, critique, and repair.
• scoville-ui-anti-ai-slop: framework UI implementation, accessibility mechanics, and rendered proof.
• scoville-scribe-anti-ai-slop: wording, terminology, meaning, and source fidelity.
• scoville-plan: durable Plans, Work Items, Decisions, and lifecycle state.
• scoville-handoff: active-work transfer.

Studio workflow

Working notes are not extra deliverables. Start with concern | owner | decision/status | evidence; expand only material open fields. Carry source authority, protected constraints and unknowns between phases. Approval never means verification.

After two coherent unsuccessful passes, reassess cause and method; continue while evidence supports progress, change approach on repeated failure. There is no one-correction cap, required cosmetic repair or render quota. A clean first render can pass after applicable checks. Unknown evidence permits safe source work but not dependent acceptance.

When concept is open, let a verified subject fact visibly shape a decision (hierarchy, type, colour, image or motion); record fact → decision. If swapping logo and nouns serves unrelated subjects unchanged, examine the mechanism. Style is a domain, not a mode. Preserve useful character; novelty is not required.

At consolidation or inspection, if a device adds unrest or its contribution is unclear, briefly compare the whole without it. Retain it when grouping, rhythm or expression visibly benefits. Keep unrelated choices fixed; this is a targeted probe, not a required test of every mark or a preference for fewer elements.

Composition: decide, bind, inspect

Apply this foundation in every Design task. Steps 1 and 2 guide exploration with provisional choices; consolidate before extending repetition. Steps 3 through 6 govern inspection and final acceptance, not advance paperwork for each rough variant. Keep content and binding constraints throughout; address visible failures when found. Early targeted checks may inform an uncertain direction. Small repairs scope work to affected regions and shared consumers, without redesigning settled work.

1. Order before coordinates. Establish entry, subordinate reads and action. Group by meaning; use proximity, similarity, enclosure, connection and continuity as potentially conflicting cues. Choose spacing by job: within-group, between-group, section, caption and edge. Give negative space a purpose (separate, frame, pace, focus); neither empty nor dense is inherently good.
2. Commit a usable system. Derive axes, columns, spans, gutters and shared values from real content, medium and incumbent systems. Place the longest required unit and representative repeated unit before multiplying them. Equal roles reuse values; variation needs a role, rhythm or optical reason. Set type roles as a coupled size/weight/measure/leading system, not isolated sizes. Grid choices can change after a fit check; no universal scale, ratio, column count or font-size count applies.
3. Discover before choosing checks. Map panels, text, artwork/icons and nested/repeated groups. Include the entry, dominant recognizable object and dense text group when present and material; inspect a title with any rule or eyebrow that frames it. For a motif joining or interrupting a drawn rule, inspect both outer junctions and internal gaps as one unit, against intended overlap or separation. Observe grouping before measuring. Cross-check editable source/layers, including visible use/image instances, not unused definitions. Enclosed controls, repeated groups and elements near frame edges each need a disposition. Include neighbours and both outside endpoints of repeated rows/columns. An empty helper queue does not close this map.
4. Bind each material relation. Use object → actual region/anchor | intent + target + basis | owner | metric + view | status. Start unverified. Basis is supplied, incumbent, measured, inferred, provisional or unknown; unknown is not zero. Preserve candidates or reasoned dispositions rather than selecting them again from memory. A DOM parent, growing group box or nearest rectangle is not automatically an enclosure. Follow every relevant nested drawn boundary. Enclosure, backdrop and bleed are distinct roles. For repeated groups check internal rhythm and both external anchors; if no common span applies, retain its visible/source basis.
5. Match proof to intent. Declare axis and reference (edge, baseline, geometric or optical centre) before measuring. A one-line button/pill/badge starts with a centring hypothesis unless its system supports another alignment. Text plus icon/arrow is one compound group. A minimum-inset pass proves neither centring nor sufficient space. Inspect effective paint, strokes, corners and glyph ink; boxes are not optical mass. Choose tolerances from scale and precision before results, in one coordinate space. Preserve target and local region through repair; a larger container cannot close the original failure. If intent changes, retain the old result and explain the supported new hypothesis.
6. Reconcile the final artifact. Pair actual whole/detail visual observations and suitable measured results for every critical relation on the same final source/render version. Detail crops must derive from that render; identify it in working evidence (use hashes when files are available). Recheck changed regions and neighbours, including artwork when text fits. Retain every candidate's final disposition; uncovered, stale or unsupported checks remain open.

Load the Composition methods index entry only when an open relationship needs a detailed discovery/measurement procedure or a diagnostic comparison, and the method's required input is available. Discovery may precede binding; measurement follows it. Small repair, mark and critique labels alone do not trigger it. Without suitable inputs, record the missing proof, not a pass.

Acceptance and repair

A preferred idea may fail in execution: judge visible type, composition and image craft, not its explanation. Keep content, execution and preference distinct; missing feedback is not agreement and independent disagreement stays visible.

Diagnose in order: content/requirement → semantic order → container/measure → grid/span → scale/density → crop/source → local optical correction. Stop at the first supported cause; repair its canonical owner and inspect consumers. Do not conceal failures through clipping, blanket centring, more whitespace, smaller type or one-column stacking. Revert or narrow regressions. Cause may be confirmed, inferred or unknown.

At acceptance, an exception needs supported intent/gain, protected communication/access floors, counterstructure, cost and a falsifying observation. Compare a fair conventional control when challenged; keep non-intervention variables constant, allowing the repair's declared dependent effects. A plausible story after the fact is not evidence. Never reclassify user-rejected overflow as intentional bleed. Source, rights, consent and production constraints cannot become aesthetic exceptions. Ordinary local exceptions stay with their specialist; Critique handles a separate multi-domain lifecycle or challenged cross-domain exception.

Route and read only what applies

Use the Direct expert index. General composition foundations never require another module. Specialists retain their geometry: packaging faces, mark contours, chart encodings, map topology, responsive pressure and fixed-media structure. Independent concerns still route: label appearance does not alone select Information, but a contradicted value-to-position relation does; Wayfinding does not settle map geometry; Mark does not reopen Brand/Concept unless open. Brand also owns an authored icon-family or stroke system outside identity work. Revisit routing when a new concern appears; no fixed module-count/token gate.

Verify received coverage before applying a selected source: numbered lines, total, received and missing intervals. Recover missing or cut-through lines, including intact adjacent lines; reconcile each file separately. Keep conflicting returns and source order. An end marker is not proof of a complete middle. Use the optional source reader or an equivalent. Its header contains recovery details when transport fails. If coverage cannot be recovered, carry the dependent concern unverified. Headings do not unload context; split phases only for actual task/capacity needs. User/provider limits and authorisation remain binding.

Proof ceiling

Distinguish binding constraints, evidence-bounded rules, conventions, heuristics, preferences and exceptions. Numeric advice needs its population, language/script, medium, task and override basis; one study or starting value is no universal law. Render proof covers inspected appearance only. It requires a host capability through which the agent actually views the rendered artifact or its image, such as an image-viewing tool, a screenshot that the agent can open, or a supplied render. Creating a screenshot file, reading its bytes or metadata, or inspecting source alone is not viewing. Without such a capability, record every visual-lane result as unverified and state that limit. Syntax, runtime, accessibility, user research, rights, provider and physical acceptance are separate lanes. Check current standards, law, licences and living-community claims when required; if unavailable, mark unverified and stop at that authority boundary.

Terms

• Module / expert / leaf: one linked reference and its domain role, not a subagent.
• Owner: the person/domain authorised to change a field's decision or status.
• Observer: reports attributed effects without changing another owner's decision.
• Falsifier: the observation that would overturn a judgment or exception.
• Control: a fair comparison preserving unrelated conditions.
• Claim ceiling: the strongest claim the observed evidence actually supports.

Expert routing

SKILL.md itself is the Core. Load expert references only through the exact links in the Direct expert index, never through a guessed Core file or transitive sibling loads. Core or an active module may also link an optional packaged tool in scripts/; read or execute that exact tool when needed. This exception does not load other experts. Core or an active module may similarly link an exact declared teaching source or render in examples/. Inspect only that named asset when it can resolve the current decision; an example is neither a hidden module nor proof of the result. modules.yaml is non-routed packaging/evaluation metadata, not another runtime router. Repository tooling generates this index from its labels and paths and checks for drift before packaging; runtime routing remains here in SKILL.md. This installed snapshot does not itself attest that those development checks ran. Neither the metadata nor the repository tools are runtime load targets. Module Sources: IDs and the non-routed source index are provenance, not load instructions. Runtime rules remain local and offline except an actually required source check.