benjaminstelzer Scoville Scribe Anti-AI-Slop
Write effective artifact text without changing truth, meaning, canonical terms, or behavior, and without filler.
Explicit opt-out: load no references; use no Skill-directed tool; change no text; make no Skill-derived claim. Report higher-authority requirements to use Scribe.
Activate only for text artifacts
If Scoville Plan is installed, available, and applicable to native Plans, Work Items, or Decisions, Plan owns their wording and fidelity. Do not activate, load, or apply Scribe to those records, including requested rewriting and wording audits, unless the user explicitly requests Scribe. This exception also applies to follow-up transformations and Scribe already active for another segment. It does not cover unrelated prose in a mixed task.
If Plan is absent, unavailable, or inactive, use Scribe's normal activation rules and preserve the existing record owner's format and lifecycle. Never require, install, or simulate Plan to do that work.
When explicitly requested for native record wording, keep fixed short fields Core-only within Scribe. If Plan is available and applicable, also load Plan and its routed references to check permitted edits, format, and lifecycle. Route independently requested continuous prose through Prose.
Activate Scribe when wording itself is a requested deliverable or transformation target: drafting, editing, rewriting, summarizing supplied source material, localizing, source-exact work, wording or fidelity audit, named-audience or named-genre copy, publication-ready or paste-ready text, explicit reuse, or reader-facing interface text.
Do not activate Scribe for ordinary answers, analysis, explanations, status or progress recaps, domain review findings, research answers, final framing, or a domain owner's normal record merely because the result contains prose. Chat, file, commit, or other delivery alone neither activates nor suppresses Scribe. Incidental comments, docstrings, commit messages, and pull-request text remain with Code; explicitly requested drafting or rewriting of those artifacts adds Scribe, with Code retained when engineering truth must be derived or verified.
Classify mixed tasks per segment. A domain owner's normal result does not add Scribe without an independent wording or fidelity trigger. A referential follow-up transformation of an active artifact reactivates Scribe; a pivot to ordinary explanation does not inherit it. Fully supplied fixed insertion with no wording, fidelity, terminology, behavior, or interface judgment does not activate Scribe unless the user explicitly invokes it.
For Source-exact extraction or reproduction, preserve the requested boundary and that span's whitespace and newline state. When markers occupy their own lines, the separator newline before the end marker is not part of the selected span unless the boundary explicitly includes it. For whole-file Source-exact or opt-out passthrough, read and write the file through a byte-preserving path that keeps encoding, line endings, and trailing whitespace intact; never through a reader that splits lines, normalizes newlines, or re-encodes text. Serialize directly. Fixed insertion copies the supplied string exactly.
Resolve authority and scope
Resolve compatible host, project, request, and Scribe constraints per segment during writing:
| Concern | Ordered authority |
|---|---|
| Current truth | Verified behavior; supplied facts |
| Target truth | Engineering evidence before completion/publication |
| Terms | Explicit user terms; glossaries; shared strings; established same-concept usage |
| Form | Explicit requirements; house style; genuine voice samples; surface/genre conventions |
Report material conflicts. Behavior outranks terms; facts outrank style. Resolve only output-affecting reader, allowed facts/claims, voice, language, genre, form, and integrity. Ask only when an unknown changes output materially.
Scribe owns wording/meaning and minimal targeted read-only term/text inspection.
Family standalone: discovery != installed|active|applicable|required;
absent|inactive => ignore/no require|install|simulate|reimplement;
active+applicable => owner concern only, self continues; opt-out local. Owners:
scoville-brainstorm divergence;
scoville-research source-backed research and synthesis;
scoville-code-anti-ai-slop engineering/proof; scoville-ui-anti-ai-slop
interface/rendered proof; scoville-plan records/lifecycle;
scoville-handoff transfer. Reuse sibling evidence; without Plan, preserve owner format/lifecycle
and edit permitted wording only. Never infer behavior from copy. Fixed
source-exact rendering does not trigger UI.
After safety and explicit constraints, serve the artifact reader's required knowledge, decision, or action. Change text only for a binding request or convention, a concrete defect, or that outcome. Match detail; include required causal links and boundaries, plus useful examples.
Select mode and references per segment
Route by the requested transformation and target artifact surface. Neither the integrity floor nor comparison against facts, requirements, behavior, or a glossary selects Fidelity; add it only for an explicit Fidelity operation below. Combine routes only when each route independently triggers for the same segment.
Selecting exact tokens from a structured requirements record is not Source-exact extraction/reproduction. A stepwise non-interface procedure stays Core-only unless a Fidelity-row trigger independently applies.
Report any requested route or profile from the selected references: none is
CORE_ONLY; otherwise report exactly the selected routes in controller order.
| Mode | Core rule |
|---|---|
| Fixed insertion | Fully supplied wording plus fixed structure, with no wording, fidelity, terminology, behavior, or interface judgment: insert and load no reference despite surface triggers. Extraction/reproduction is Source-exact. |
| Structured grouping | Retain every enumerated item and unresolved lifecycle choice; load no reference. |
| Draft | Use only permitted facts/claims. |
| Edit | Make the smallest sufficient change. Claim-preserving rewrite is Edit. |
| Audit | Report location, problem, reader effect, correction direction. Rewrite only if asked; follow the subject's route. |
Reference routes:
| Read | Before or for |
|---|---|
| Fidelity | Adaptation; material summary; localization; Source-exact extraction/reproduction; audits of those outputs; controlled variants; regulated or author-owned drafting/editing. |
| Interface | Writing, changing, auditing, or comparing any non-Fixed-insertion user-facing GUI/CLI string, error, notification, transactional message, accessible text, metadata, or behavior-bound procedure. Product-generated email uses Interface. Editorial, personal, and newsletter prose follows prose routing unless behavior-bound. User-facing localization also needs Fidelity. |
| Prose | Continuous artifact prose; independently requested substantive rationale; requested natural, human-sounding, or anti-slop artifact writing. Use Prose instead of Fidelity for continuous claim-preserving rewrites unless Fidelity also triggers. Exclude domain-owned normal records, format-owned fixed short records, and behavior-bound interface procedures. A non-interface artifact procedure triggers Prose only when its explanatory wording is independently variable. Audit artifact prose claims with Prose unless Fidelity triggers. |
Classify mixed artifacts only far enough to route each segment correctly.
Preserve integrity
Never invent or silently change: names; numbers, dates, units; links, citations; attribution, quotations, exact-text boundaries; schemas, technical terms; negation, modality, conditions, exceptions; first-person experience, identity, relationships, opinions, feelings; product capabilities, causes, guarantees, timelines, actions, or verified runtime behavior. Mark hypotheses/examples as hypothetical. Preserve unresolved deliberate ambiguity.
Behavior-bound text states the supported current state. State a requested target only in its artifact. Preserve procedure prerequisites, order, inputs, warnings, commands, placeholders, and expected results unless request or verified behavior changes them.
High-risk legal, medical, financial, safety, privacy, or publication claims need appropriate authoritative sources and qualified review; editing is neither.
Preserve strong passages and demonstrated register. Never diagnose AI authorship, promise detector evasion, or vary canonical terms for style. Beyond explicit constraints and the punctuation house style below, do not invent universal word, punctuation, sentence-length, or formatting bans. Apply routine guidance silently. Disclose only host-required non-obvious actions, pauses, scope changes, external effects, or material risk.
Punctuation house style
Build newly written or edited natural-language text, including interface copy,
around complete sentences and grammatical clauses, primarily using periods
and commas. Do not use em dashes (U+2014), en dashes (U+2013), or semicolons (;).
Rewrite affected sentences, not just their punctuation. Integrate asides into the main clause, connect related ideas grammatically, or split a thought into self-contained sentences while preserving its logical relationships. Do not keep the old construction by swapping in other marks, including hyphens, colons, parentheses, commas, or periods.
If a sentence genuinely needs a dash, use only the ASCII hyphen (-) and use
it sparingly. Keep necessary compound-word hyphens and required list markers.
Preserve punctuation inside exact quotations, Source-exact or fixed-insertion content, code, commands, URLs, identifiers, and canonical names. These rules govern newly written prose and prose within the requested edit scope, not protected source text or syntax. Leave existing text outside that scope unchanged. The rules apply across languages and do not imply that punctuation proves AI authorship.
Complete
Before return, verify reader outcome, touched integrity items, each segment's surface/fidelity route, requested language/format/length/mode, sentence-first punctuation house style within the requested writing or edit scope, and removal of unrequested process commentary. Report unresolved factual/project conflicts. Audit-only means no edit.