Slop Curator

View saved version →

---
name: arabic-style-curator
description: Curate Arabic prose for dialect authenticity, natural Arabic-English code-switching, house vocabulary, Egyptian spoken cadence, banned AI expressions, and phrase-bank guidance. Use for Egyptian Arabic, Saudi Arabic, MSA, neutral Arabic, bilingual content, or when Arabic sounds translated, generic, or mechanically formal.
---

# Arabic Style Curator

Own Arabic language behavior, not the whole editorial workflow.

Use this Skill after meaning and factual locks are clear. Pair it with `plain-spoken-writing` when sayability and spoken thought order matter.

## First decide the Arabic mode

Choose exactly one primary mode unless the user explicitly requests a mix:

- Egyptian Arabic
- Saudi Arabic
- Modern Standard Arabic
- neutral regional Arabic
- bilingual Arabic-English

Do not mix dialects accidentally.

## Core rules

- write Arabic in natural Arabic thought order, not translated English syntax
- preserve standard industry English when the speaker would actually use it
- do not insert English merely to sound knowledgeable
- maintain agreement, pronouns, gender, number, and tense
- keep Arabic punctuation restrained
- preserve Latin product and brand names in their original direction
- keep terminology consistent across Arabic and English

## Egyptian and spoken mode

When Egyptian Arabic or spoken copy is requested, load:

- `references/egyptian-house-voice.md`
- the full `plain-spoken-writing` Skill

Use the phrase bank as **rhythm evidence**, never as copy-paste inventory. Repeatedly recycling the same opener, check-in, or transition is itself slop.

## AI-expression checks

Load:

- `references/ai-expression-bank.md`
- `references/english-ai-expression-bank.md` for bilingual work

Under the portable profile, treat these as strong signals whose context still matters. Under the strict house-style profile, treat the listed expressions as hard bans unless exact reproduction requires them.

## House vocabulary

When the strict house profile is active, enforce the banned Arabic vocabulary in `strict-human-output/references/house-style.md`.

Do not evade a banned word with a near-identical metaphor. Rewrite the underlying idea literally.

## Connection with Plain-Spoken Writing

`plain-spoken-writing` owns sayability, thought order, and personal cadence.

`arabic-style-curator` owns dialect consistency, Arabic syntax, terminology, Arabic AI-expression detection, and house vocabulary.

When they disagree, preserve factual meaning first, then the speaker's demonstrated voice, then the active house profile.

## Completion condition

The Arabic should read like native writing for the requested market, preserve the intended voice, keep professional English only where natural, and pass any active house bans without sounding mechanically sanitized.

Referenced files: 4

View saved version →

---
name: commercial-copy-director
description: Build or edit advertising copy, landing-page copy, campaign copy, marketing playbooks, offers, headlines, and CTAs using audience awareness, real objections, proof, specificity, and decision context. Use when the writing must persuade commercially without fabricated proof, fake urgency, generic hype, or psychological pressure.
---

# Commercial Copy Director

Own the commercial job of the copy. Do not use persuasion language as decoration.

## Before writing

Silently identify from available context:

- audience
- awareness level
- immediate job to be done
- triggering situation
- primary desired outcome
- strongest real obstacle
- largest perceived risk
- main objection
- available proof
- offer
- requested action

If one detail is missing but the task remains safe and useful, use the smallest reasonable assumption or a visible placeholder. Do not invent proof or operational facts.

## One primary message

Choose one main message for each piece of copy. Support it with concrete details, use cases, inputs, outputs, limitations, decisions, timing, and proof.

Do not jump from a feature to an emotional promise without explaining the practical connection.

## Ethical persuasion constraints

Never invent or exaggerate:

- statistics, case studies, customer quotes, testimonials, awards, rankings, capabilities, or performance outcomes
- scarcity, urgency, deadlines, limited seats, high demand, or price changes
- audience insecurity, shame, status anxiety, or failure as the sole reason to buy

Do not insult the audience or competitors.

## Headlines

A useful headline should communicate at least one concrete element:

- audience
- problem
- desired result
- use case
- product category
- offer
- proof
- time
- constraint
- specific action

Avoid pure curiosity, dramatic declarations, total-change promises, and repeated formulas across a headline set.

## CTAs

Describe the next real action, such as:

`حمّل الدليل، اطلب عرض السعر، احجز موعدًا، ابدأ التجربة، شاهد تفاصيل الباقة، أرسل بيانات المشروع، قارن الخيارات، راجع المحتويات`

or

`Download the guide, View the package, Request pricing, Book a call, Start the trial`

Avoid empty aspiration CTAs such as journey, leap, potential, future, transformation, or generic discovery language when a more specific action exists.

## Playbooks

For practical playbooks:

- organize around decisions and recurring situations, not generic theory
- include actionable frameworks only when they serve a real decision
- use examples as examples, never disguised case studies
- distinguish rules, diagnostics, examples, templates, tools, time, inputs, outputs, and failure modes
- do not force every chapter into the same shape
- preserve practical English terminology when the audience uses it

## Routed specialists

- use `arabic-style-curator` for Arabic/dialect/bilingual copy
- use `plain-spoken-writing` when the copy should sound spoken
- use `strict-human-output` for strict house rules and evidence discipline
- use `slop-quality-gate` before final delivery

## Completion condition

The copy should make the intended action easier to decide based on clear, truthful reasons, with no invented proof and no generic language standing in for a real mechanism or offer.

Referenced files: 2

View saved version →

---
name: plain-spoken-writing
description: Use when drafting or rewriting prose that should sound like the speaker actually talks, especially Egyptian Arabic, Arabic-English code-switching, social posts, emails, business writing, personal voice, conversational copy, or when output feels polished, generic, formal, or AI-written.
---

# Plain-Spoken Writing

## Core principle

**Write like the speaker speaks, not like a writer pretending to sound conversational**

Start from spoken thought. Edit only enough to make it easy to read

Target **clean speech**, not polished prose and not a raw transcript

Priority when rules compete:
1. meaning and factual accuracy
2. the speaker's real voice
3. natural thought order
4. readability
5. polish

Polish is last on purpose

## Input mode

**If a speech sample, transcript, rough note, or prior writing exists:** treat it as the strongest style evidence. Preserve vocabulary, sentence movement, code-switching, useful repetition, direct address, rhetorical questions, side comments, and formality

**If no voice sample exists:** explain the idea as if speaking to one real person, then clean lightly. Never invent slang, quirks, opinions, stories, clients, meetings, or experiences to simulate humanity

## Workflow

### 1. Hear the sentence first

Silently ask: **How would this person actually say this out loud?**

Do not optimize for the most elegant written sentence

### 2. Keep useful roughness

Keep roughness when it carries identity or timing: a sentence that circles the point, a quick correction, deliberate repetition, a short aside, a question answered immediately, uneven paragraph lengths, or natural professional English terms inside Arabic

### 3. Remove noise, not personality

Clean accidental duplicated meaning, useless filler, broken transcription, ASR artifacts, and false starts that block understanding

Do not mechanically delete repetition or dialect

### 4. Make speech readable

Add only what reading needs: paragraph breaks, enough punctuation, corrected distracting errors, and minor clarification

Do not turn the piece into an essay

### 5. Run the voice test

Read the draft mentally aloud. If the tongue naturally wants to replace a sentence, use the spoken version

**If you wouldn't say it, don't write it**

### 6. Humanity pass

Before delivering:
- cut repeated ideas, not repeated emphasis
- replace generic professional wording with the speaker's normal vocabulary
- remove perfect transitions, quote-card wisdom, and artificial engagement bait
- break suspiciously uniform paragraph rhythm
- verify no personal detail was invented
- keep any promotion in the same speaking voice

For scoring and hard-fail checks, read [references/evaluator.md](references/evaluator.md)

## Arabic-English code-switching

Do not translate normal professional terms merely because the surrounding text is Arabic

If the speaker naturally says `CRO`, `prompt`, `workflow`, `skill`, `plugin`, `marketplace`, `output`, or `customer research`, keep them when appropriate

Do not add English just to sound modern

For Egyptian-Arabic handling, read [references/egyptian-arabic.md](references/egyptian-arabic.md)

## Structure

Follow the thought's movement, not a default content framework

Do not automatically impose Hook → Problem → Lesson → CTA, introduction → three points → conclusion, repeated `مش X، لكن Y`, or perfectly symmetric paragraphs

A post can still have a hook or CTA when the task needs one. They should emerge naturally rather than expose the template

## Explain instead of defining

Prefer spoken explanation over dictionary-style definition

Too written:
> A marketplace is a repository containing multiple reusable capabilities

Plain-spoken:
> طب الـ marketplace ده إيه أصلًا؟ أنت داخل مكان واحد، جواه شوية Skills وAgents وحاجات متجمعة حوالين نوع شغل معين

## Repetition

**Keep** repetition carrying emotion, emphasis, timing, or personality

**Cut** repetition that only restates the same information

## Never fake humanity

Human does not mean random slang, intentional typos, fake uncertainty, or invented anecdotes

If the user supplied an experience or claim, it may be used. Otherwise do not manufacture it

## Promotion

When mentioning a product or project, connect it to the thought already being discussed. Do not switch into announcement or launch-copy voice

For posts, emails, explainers, and business writing, read [references/writing-modes.md](references/writing-modes.md)

## When the draft still feels AI-written

Read [references/anti-ai-patterns.md](references/anti-ai-patterns.md)

Typical symptoms: perfect transitions, generic insight sentences, repeated contrast formulas, fake slang, unnecessary summaries, motivational conclusions, and generic CTAs

## Source voice outranks examples

If the user provides samples, derive the local voice from them rather than copying phrases from this skill

For a reusable profile across multiple samples, use [assets/voice-profile-template.md](assets/voice-profile-template.md)

## Output contract

Unless critique or alternatives are requested, deliver the writing directly. Do not explain the technique, label hook/tension/CTA, or append a quality report

When rewriting, preserve meaning and claims unless explicitly asked to change them

## Final five checks

1. **Sayability:** could this speaker actually say it aloud?
2. **Identity:** does the vocabulary belong to this speaker?
3. **Movement:** does the thought move naturally rather than expose a template?
4. **Cleanliness:** is it readable without becoming article prose?
5. **Truthfulness:** was no fake human detail added?

If one fails, revise

Referenced files: 10

View saved version →

---
name: slop-audit
description: Audit prose for named AI-slop patterns without rewriting or guessing authorship. Use when the user asks what sounds generic, formulaic, robotic, over-polished, or AI-like and wants evidence instead of an edit.
---

# Slop Audit

Inspect the draft as evidence. Do not rewrite it unless the user later asks for an edit.

## Method

1. Read the full passage before flagging anything.
2. Identify genre and likely voice conventions from the text itself.
3. Match only named patterns in `references/pattern-taxonomy.md`.
4. Classify each finding as `high`, `contextual`, or `uncertain` confidence.
5. Quote the smallest useful span, name the pattern, explain why it weakens this passage, and give a short repair direction.
6. Merge duplicate findings when one line triggers several labels for the same underlying problem.
7. Do not count a stylistic choice as slop merely because it contains an adverb, passive voice, a question, a fragment, a Wh-word opening, or an em dash.
8. Do not estimate the probability that AI wrote the passage.

## Output

Use this compact shape:

- **Pattern:** name
- **Evidence:** short quote
- **Confidence:** high / contextual / uncertain
- **Why it reads generic here:** one sentence
- **Fix direction:** a few words

Then add `Overall` with the 2 or 3 dominant tendencies. Do not produce a fake precision score unless the user asks for one.

## Optional static signal check

For long plain-text or Markdown drafts, `scripts/short_line_lint.py` can flag clusters of unusually short standalone prose lines. Treat its output as contextual evidence only. Ignore matches that are clearly poetry, lyrics, dialogue, UI copy, headings, lists, tables, code, or an explicitly requested line-by-line format.

## No-finding behavior

If the passage is already clean, say so. Do not manufacture findings to justify the Skill.

Referenced files: 3

View saved version →

---
name: slop-pattern-repair
description: Repair specific recurring phrase, structure, rhythm, agency, metadiscourse, and formatting patterns after a draft has been read in context. Use when the user wants targeted anti-slop cleanup rather than a general rewrite.
---

# Slop Pattern Repair

Repair named patterns, not the writer's personality.

## Procedure

1. Read the whole draft first.
2. Identify only the regions with a clear pattern problem.
3. Use `references/repair-playbook.md` to choose the smallest repair.
4. Preserve factual locks and voice anchors from the surrounding prose.
5. Re-read the paragraph after each repair. A local fix can damage cadence or meaning even when the sentence looks cleaner alone.
6. Do not normalize untouched sentences merely to match the edited ones.

## Rule weighting

### Usually repair

- throat-clearing before the point
- formulaic binary reframes
- repeated negative reveals
- faux-insight framing
- unsupported authority
- importance labels without consequences
- repeated metacommentary
- synonym cycling
- fake-profound endings
- formatting that exists only for emphasis

### Repair only in context

- adverbs
- passive voice
- inanimate agency
- rhetorical questions
- Wh-word openings
- fragments
- list length
- em dashes
- paragraph length
- short-line stacking or micro-paragraphing
- contractions or informality

A contextual pattern deserves a repair when it clusters, hides meaning, repeats a template, or obviously conflicts with the genre. One instance is often just writing.

## Repair style

Prefer deletion before replacement when the sentence still works without the slop. When replacement is necessary, use the draft's own vocabulary and syntax as the source of style.

## Spoken-language veto

When `plain-spoken-writing` is in the route, do not remove a dialect marker, filler, rhetorical question, direct address, correction, code-switched term, long spoken sentence, or repetition merely because a generic anti-slop rule can match it. Repair only when the feature is redundant, templated, unclear, or demonstrably weakens this specific piece. If a repair makes the text less sayable or more like article prose, return that region to `plain-spoken-writing`.

Referenced files: 2

View saved version →

---
name: slop-quality-gate
description: Check an edited or newly written draft for factual drift, voice loss, over-editing, residual formulaic patterns, robotic rhythm, sayability failures, and output-contract failures. Use after a rewrite, a plain-spoken pass, or when the user wants a quality check.
---

# Slop Quality Gate

Evaluate the finished draft independently. Do not reward the edit merely for being shorter or more polished.

## Hard gates

Fail immediately if the edit:

- invents or changes a fact, number, date, quote, source, name, or claim
- asserts that AI wrote the text
- changes the user's intended position
- removes a meaningful constraint or qualification
- flattens a distinctive voice into generic professional prose
- turns requested spoken or dialect writing into formal article prose
- fabricates slang, personal stories, uncertainty, or other "human" details
- invents proof, scarcity, urgency, testimonials, rankings, capabilities, or unsupported precision
- violates an explicitly active house-style hard rule

## Scorecard

Score internally out of 100:

- factual and semantic fidelity: 25
- voice retention: 20
- specificity and concrete support: 15
- natural rhythm: 15
- directness: 10
- structural variety without theatrics: 10
- formatting restraint: 5

When `commercial-copy-director` was used, evidence integrity and truthful offer framing are hard requirements regardless of the numeric score. When `strict-human-output` house mode was active, active hard bans must pass.

When `plain-spoken-writing` was part of the route, treat sayability, natural thought order, dialect authenticity, and code-switching fidelity as part of voice retention and natural rhythm. Do not reward grammatical polish that makes the result less speakable.

Default pass threshold: 85, with no hard-gate failure.

Do not show the numeric score unless the user asks for scoring.

## Residual checks

Check for:

- repeated binary reframes
- repeated throat-clearing or faux-insight setups
- unsupported authority
- vague importance claims
- synonym cycling
- fake-profound ending
- repeated question-answer scaffolds
- stacked punchy fragments
- excessive short-line stacking or micro-paragraphing
- uniform paragraph geometry
- formatting used only for emphasis
- second-order anti-slop artifacts such as forced slang or over-compression

## Feedback routing

If the draft fails, return only the failing regions to the responsible pass:

- meaning, evidence, factual locks, or general existing-writer voice -> `voice-preserving-edit`
- sayability, spoken thought order, Egyptian dialect, natural code-switching, or controlled roughness -> `plain-spoken-writing`
- phrase, structure, rhythm, agency, or formatting -> `slop-pattern-repair`
- Arabic dialect, Arabic syntax, code-switch terminology, or house Arabic vocabulary -> `arabic-style-curator`
- active hard policy, evidence integrity, format compliance, or punctuation contract -> `strict-human-output`
- headline, CTA, offer framing, proof, or commercial-decision failure -> `commercial-copy-director`
- presentation/document hierarchy, fake visual content, RTL design, or asset-control failure -> `visual-content-anti-slop`

Recheck the revised regions and then the full piece once. Avoid endless polishing loops.

## Output behavior

When invoked as part of another workflow, keep the gate internal and return the finished draft. When the user explicitly asks for a quality review, return pass/fail by category with concise evidence.

Referenced files: 1

View saved version →

---
name: slop-router
description: Route a writing or content request through the minimum connected Slop Curator Skills needed. Use for anti-slop editing, voice preservation, spoken or Egyptian Arabic writing, strict house-style rules, commercial copy, playbooks, presentations, visual-content review, or combined audit-and-edit workflows.
---

# Slop Router

Act as the front door for Slop Curator. The Plugin contains multiple full Skills. Route deliberately rather than loading every rule pack into every request.

## Global priority order

1. exact approved text when exact reproduction is requested
2. factual, legal, numeric, source, and technical fidelity
3. current user intent and explicit constraints
4. demonstrated writer/brand voice
5. active language, dialect, house-style, and format profile
6. genre and audience conventions
7. anti-slop cleanup and stylistic improvement

A lower priority never silently overrides a higher one.

## Classify the job first

### Evidence-only audit

Route: `slop-audit`

Do not rewrite or guess authorship.

### Standard anti-slop edit

Route: `voice-preserving-edit -> slop-pattern-repair -> slop-quality-gate`

Add `strict-human-output` portable baseline when factual claims, format constraints, or explicit strictness matter.

### Spoken or conversational writing

Route: `plain-spoken-writing -> optional slop-pattern-repair -> slop-quality-gate`

For Egyptian Arabic or bilingual Arabic-English work, insert `arabic-style-curator` before the gate.

### Preserve my voice + make it spoken

Route: `voice-preserving-edit -> plain-spoken-writing -> language specialist if needed -> optional repair -> slop-quality-gate`

### Strict house-style / my writing rules

Route: `content owner -> strict-human-output(house) -> slop-quality-gate`

If Arabic is involved, insert `arabic-style-curator`. The house profile includes hard punctuation, banned-vocabulary, and formula constraints. Do not apply it to unrelated public users without an activation signal.

### Advertising / campaign / landing page / marketing copy

Route: `commercial-copy-director -> language/voice specialist as needed -> strict-human-output(portable or house) -> slop-quality-gate`

Commercial copy owns the audience, awareness, objection, proof, offer, headline, and CTA decisions.

### Practical marketing playbook

Route: `commercial-copy-director -> arabic-style-curator or plain-spoken-writing as requested -> strict-human-output(house when active) -> slop-quality-gate`

### Presentation / report / brochure / visual brief / RTL content

Route: `visual-content-anti-slop -> arabic-style-curator if Arabic -> strict-human-output if active -> slop-quality-gate`

This route governs content and art-direction constraints. Use actual visual/file tools only when the current host provides them.

### Audit then edit

Run `slop-audit` first. Convert its evidence into the smallest repair plan, then use the appropriate content owner and final gate.

## Rule-pack selection

Do not confuse universal rules with personal preferences.

- `portable baseline`: evidence integrity, specificity, no fabricated proof, useful structure, user constraints
- `house profile`: terminal-period ban, em-dash ban, hard phrase/vocabulary bans, response behavior, personal punctuation conventions
- `Arabic profile`: dialect, Arabic syntax, code-switching, Arabic AI-expression bank, house phrase bank
- `commercial profile`: audience, awareness, offer, proof, headline, CTA, playbook behavior
- `visual profile`: document hierarchy, real assets, verified charts, RTL design, anti-decoration rules

Load only the profiles needed for the request.

## Neural connection rules

The graph in `references/connection-map.json` is the machine-readable source of truth.

Important feedback behavior:

- voice flattening -> `voice-preserving-edit`
- spoken/dialect sanitization -> `plain-spoken-writing`
- Arabic syntax, code-switching, or house Arabic failure -> `arabic-style-curator`
- phrase/structure/rhythm residue -> `slop-pattern-repair`
- hard policy, evidence, punctuation, or exact-format failure -> `strict-human-output`
- weak commercial decision logic, headline, CTA, proof, or offer framing -> `commercial-copy-director`
- visual hierarchy, fake content, RTL rendering logic, or asset misuse -> `visual-content-anti-slop`
- all routes converge on `slop-quality-gate` when a final review is appropriate

## Signal weighting

Use three levels:

- **hard:** fabricated evidence, broken user constraint, active house ban, formulaic claim with no useful context
- **contextual:** adverbs, passive voice, rhetorical questions, fragments, Wh-openers, em dashes outside house mode, list length, paragraph length
- **voice:** dialect markers, code-switching, useful repetition, asides, corrections, direct address, controlled roughness

Do not turn contextual signals into universal bans.

## Finish

Return the requested deliverable in the requested format. Keep internal routing and quality checks invisible unless the user asks for them.

Referenced files: 6

View saved version →

---
name: strict-human-output
description: Apply a structured human-writing policy to prose when the user wants strict anti-slop constraints, house-style compliance, factual discipline, punctuation controls, banned-language checks, or a final rule-based pass without changing approved meaning.
---

# Strict Human Output

Use this Skill as a **policy overlay**, not as a replacement writer. Its job is to enforce explicit output constraints after the content owner has established meaning, evidence, voice, and genre.

## Two profiles

### Portable baseline

Use by default when this Skill is selected. Enforce rules that generalize well across users:

- do not invent facts, numbers, proof, testimonials, sources, rankings, capabilities, scarcity, or urgency
- remove unsupported importance claims, generic hype, theatrical transitions, and empty adjectives
- prefer specific actions, concrete nouns, observable behavior, and supported consequences
- preserve approved names, dates, prices, figures, claims, quotations, terminology, and technical syntax
- do not restate the brief, add filler introductions, or append generic closing offers
- follow requested language, dialect, length, format, order, assets, and exact-content constraints
- when a style rule conflicts with factual or technical accuracy, accuracy wins

Read `references/portable-baseline.md` for the detailed baseline.

### House-style profile

Activate only when the user explicitly requests the strict house style, says to use their personal rules, provides an active style profile with these constraints, or the surrounding task context already establishes them.

This profile adds hard constraints such as:

- no terminal full stop in normal prose lines
- no em dash
- no negative-setup + positive-reveal formulas
- no "difference" formulas or banned house vocabulary
- no copied AI-sounding Arabic/English expressions
- reduced quotation marks and decorative punctuation
- exact response-start and response-end behavior

Read `references/house-style.md` before enforcing this profile.

Do **not** impose the house profile on an unrelated public user merely because the Plugin contains it.

## Conflict order

When rules collide, use this order:

1. exact reproduction explicitly requested by the user
2. factual, legal, technical, numeric, and source fidelity
3. current user instruction
4. active brand or house-style profile
5. genre and audience conventions
6. generic anti-slop preferences

A lower rule must never silently damage a higher one.

## Workflow

1. Identify which profile is active: `portable` or `house`
2. Lock facts, exact text, technical syntax, and mandatory content
3. Scan for hard violations only
4. Rewrite the smallest necessary region rather than normalizing the whole piece
5. Recheck natural reading after every hard-rule repair
6. Hand residual voice or rhythm problems back to the relevant writing Skill
7. Run `slop-quality-gate` after enforcement when this Skill is part of a routed workflow

## Do not overcorrect

A strict style policy can itself create mechanical prose. Do not:

- remove natural uncertainty merely because it looks less decisive
- flatten dialect or code-switching
- replace every adjective with a long explanation
- force active voice when passive voice is clearer or the actor is genuinely unknown
- remove punctuation required for code, URLs, email addresses, decimals, versions, file names, citations, or exact quotations
- manufacture slang to compensate for formal prose

## Deterministic lint

When the host can run Python and the strict house profile is active, `scripts/strict_human_lint.py` can mechanically check terminal periods, em dashes, selected banned vocabulary, selected AI-expression families, and negative-setup formulas. Treat its findings as candidate violations to repair in context. It does not judge voice, factual accuracy, persuasion quality, or whether a phrase is quoted/required.

## Completion condition

The output is complete when every active hard rule passes **and** the text still preserves the intended meaning, evidence, voice, genre, and technical correctness.

Referenced files: 5

View saved version →

---
name: visual-content-anti-slop
description: Review or plan presentations, reports, brochures, documents, visual briefs, and Arabic RTL layouts for AI-design clichés, fake data, decorative copy, asset misuse, weak hierarchy, and Arabic rendering risks. Use when the requested deliverable has a visual or document-design component.
---

# Visual Content Anti-Slop

This Skill governs **content and art-direction constraints** for visual deliverables. It does not claim the host can render or edit an image unless a compatible tool is actually available.

## One job per page or slide

For presentations, reports, proposals, brochures, and client files:

- give each page or slide one primary communication job
- build a clear reading sequence
- preserve approved meaning and exact facts
- do not repeat the title as body copy
- do not add decorative copy to fill empty space
- do not invent quotations, case studies, charts, statistics, screenshots, or portfolio work
- use tables only when comparison is actually needed
- keep labels short and body text readable
- label genuine placeholders clearly

## Anti-slop visual filter

Do not default to generic AI-decoration patterns such as:

- random gradients, neon, purple glow, glassmorphism, floating translucent cards
- excessive rounding, shadows, icons, pills, badges, section numbers, or type sizes
- random 3D/chrome/liquid objects, abstract blobs, decorative waves, unexplained dots or lines
- robots, glowing brains, circuit-filled heads, futuristic tunnels, space, cyberpunk
- fake dashboards, fake app screens, fake analytics, fake charts
- stock teams pointing at screens, handshakes, rockets, light bulbs, targets, puzzle pieces, chess pieces, labyrinths, megaphones, magnifying glasses
- floating social logos, purposeless laptop/device mockups, oversized quotation marks, decorative English labels
- arbitrary diagonal layouts, busy-image text overlays, center-aligned long paragraphs

Every visual element should support comprehension, hierarchy, brand recognition, navigation, comparison, or explanation. Whitespace is allowed.

## Structure

Prefer a clear grid, consistent margins, controlled spacing, hierarchy, readable typography, sufficient contrast, purposeful whitespace, consistent alignment, limited color roles, limited type roles, and real supplied assets.

Use diagrams only when relationships need explanation, charts only when verified data exists, mockups only for real application/context, photography only when relevant, and icons only when they improve recognition or scanning.

## Arabic visual work

For Arabic presentations, graphics, documents, interfaces, or visuals:

- use true RTL layout
- use fonts that support Arabic shaping
- render connected Arabic letters correctly
- never reverse, mirror, or separate Arabic letters
- verify similar forms and hamza/ligature behavior
- check numeral and punctuation direction
- inspect mixed Arabic-English lines
- keep Latin brand names and logos in their original direction

If the available tool cannot guarantee correct Arabic rendering, state that limitation rather than pretending the render is correct.

## Asset control

When the user supplies real assets:

- use approved assets
- do not replace real assets with generated substitutes
- do not alter logos, packaging, or approved artwork unless requested
- do not add watermarked stock images or fabricate event/client photos
- if an asset is missing, use a clearly labelled neutral placeholder or redesign around the gap

## Routed specialists

Use `strict-human-output` for exact format and house-style constraints, `arabic-style-curator` for Arabic copy and RTL-language logic, and `slop-quality-gate` for final content fidelity.

Referenced files: 1

View saved version →

---
name: voice-preserving-edit
description: Edit a draft while protecting the writer's meaning, facts, cadence, vocabulary, humor, bluntness, uncertainty, and useful imperfections. Use when preserving personal voice matters more than making every sentence uniformly polished.
---

# Voice-Preserving Edit

Own meaning and voice. Anti-slop cleanup is subordinate to these two responsibilities.

## Before editing

Build a private `voice fingerprint` from the draft:

- vocabulary level and favorite kinds of words
- sentence length range and cadence
- directness versus digression
- humor, bluntness, warmth, skepticism, or formality
- use of fragments, asides, repetition, slang, profanity, or uncertainty

Also lock anything the edit must not silently change:

- names
- numbers and dates
- quotes
- claims and evidence
- product or technical terminology
- instructions, legal wording, or operational constraints

## Editing contract

1. Make the minimum effective change for the user's requested intensity.
2. Leave strong human sentences alone even if another version could sound cleaner.
3. Cut filler when it adds no context, character, or meaning.
4. Make vague claims more concrete only from information already present. Never manufacture specificity.
5. Prefer direct verbs and named actors when that improves clarity, but do not force every sentence into actor-first syntax.
6. Preserve useful roughness. A draft does not need uniform paragraph length, identical polish, or perfect symmetry.
7. Keep uncertainty when the writer is genuinely uncertain.
8. Keep genre-appropriate rhetorical devices when they have a job.
9. Reorganize only when the current order hides the point or makes the piece hard to follow.
10. If a proposed cleanup erases a voice marker, either keep the original or find a smaller fix.

## Handoff to Plain-Spoken Writing

When the user explicitly wants the result to sound spoken, conversational, Egyptian Arabic, or like something they would say aloud, keep ownership of factual locks and the voice fingerprint, then hand those anchors to `plain-spoken-writing`. Do not independently formalize or compress the draft before that handoff.

## Intensity

### Light
Correct obvious slop, repetition, and confusion. Preserve structure and most wording.

### Standard
Rewrite weak sentences, remove recurring formulas, and improve structure where needed while keeping the recognizable writer.

### Strong
Allow paragraph and section reconstruction. Keep factual locks and reintroduce the writer's vocabulary, pace, and edge before finishing.

## Output

Return the full edited draft. Unless the user requests copy only, add a short `What changed` section focused on substantive decisions, not a line-by-line change log.

Referenced files: 2