justin/no_ai_slop_writing_rules
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial commit: no-AI-slop writing rules and Rossmann voice skills

Browse Source 
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Louis Rossmann
2026-05-24 19:28:13 -05:00
commit ce03e044f3
6 changed files with 780 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.claude/skills/no-ai-slop/SKILL.md
+105
View File
@@ -0,0 +1,105 @@
---
name: no-ai-slop
description: "Rules and worked examples for writing prose that does not read like AI-generated slop. Consult before writing or editing any prose."
---
# No AI Slop
The full rule list lives in the project `CLAUDE.md` (rules 1 through 24). This skill turns the rules that have worked examples into actionable guidance: each shows a WRONG version (the slop) and a RIGHT version (the fix). The pattern behind every fix is the same: replace the vague claim with a specific, checkable fact.
## Rule 1: No emdashes
The character is banned. Use a semicolon, a period, a comma, or restructure.
- WRONG: "The policy -- which affected millions -- was later reversed."
- RIGHT: "The policy affected millions of devices. The company reversed it in December 2017."
## Rule 4: No intensifiers
"Significantly", "dramatically", "extremely" and their kin are placeholders for evidence. Replace the word with the number it was standing in for.
- WRONG: "The pricing was significantly higher than the cost of the part."
- RIGHT: "They charged $1,200 for a repair that needed a $5 chip."
## Rule 5: No hollow statements
A sentence that asserts importance without a detail says nothing. End every claim on a concrete fact.
- WRONG: "This practice has had a significant impact on people."
- RIGHT: "The company replaced 11 million batteries in 2018, against the 1 to 2 million it had expected."
## Rule 7: No structural slop (repetitive layouts)
Three sections built from the same template read as machine output, even when each fact is true. Vary paragraph count, sentence rhythm, and how each section opens.
- WRONG (three sections, identical shape):
```
In [year], [party] did [thing]. This affected [number] people. [Party] responded by [action].
In [year], [party] did [thing]. This affected [number] people. [Party] responded by [action].
In [year], [party] did [thing]. This affected [number] people. [Party] responded by [action].
```
- RIGHT (vary the shape):
```
Section one: a detailed narrative with timeline and context across two paragraphs.
Section two: a two-sentence summary, because the event is thinly documented.
Section three: opens with the party's stated justification, then the contradicting evidence.
```
## Rule 11: No filler phrases
"In today's world", "It's important to note", "When it comes to" add length, not meaning. Open on the fact.
- WRONG: "In today's world, planned obsolescence affects many devices."
- RIGHT: "Apple, Samsung, and Google have each faced lawsuits alleging planned obsolescence."
## Rule 13: Write like a researcher, not a copywriter
If a sentence could sit on any advocacy or marketing site without changing a word, it is generic. Anchor it to something checkable.
- WRONG: "People deserve the right to repair their own devices."
- RIGHT: "The FTC voted 5-0 in July 2021 to step up enforcement against illegal repair restrictions."
## Rule 15: No weasel words
"May potentially", "can help to", "might be able to" hedge a claim into meaninglessness. Either the thing happens or it does not. Say which.
- WRONG: "Serialization may potentially prevent independent repair."
- RIGHT: "Replacing an iPhone 15 camera module without the manufacturer's calibration software disables optical image stabilization."
## Rule 16: No dramatic headings
A heading names what the section holds. It does not tease, dramatize, or abstract.
- WRONG: "The Hidden Cost of Planned Obsolescence"
- RIGHT: "Economic impact of shortened product lifespans"
## Rule 19: No fabricated attributions
Never put a position in a named person's mouth from inference. State only what they actually did or said, with the real source.
- WRONG: "Senator Smith has argued that the right to repair is essential."
- RIGHT: "Senator Smith co-sponsored the Fair Repair Act in January 2024."
## Root-cause differentiation
When you contrast two things, name the concrete difference that separates them. Do not assert that one is exempt, newer, better, or unaffected without saying what specifically makes it so.
- WRONG: "2020+ Leaf models are unaffected and use the MyNISSAN app instead."
- RIGHT: "2020+ Leaf models shipped with 4G/LTE telematics units connected to a newer cloud platform, replacing the 2G/3G units in earlier models. Those vehicles use the MyNISSAN app, which talks to a different backend."
Whenever you say A differs from B, name the part, the version, the date, the mechanism, or the supply-chain change that makes the difference real. If you do not have that detail, do not imply the difference exists.
## Self-check before returning text
Run this pass on every piece of prose before you hand it back. The full banned lists are in `references/ai-writing-detection.md`; check against them directly.
1. Search for the emdash character. Remove every one (Rule 1).
2. Scan for banned verbs (delve, leverage, utilize, foster, bolster, underscore, unveil, streamline) and replace with plain equivalents.
3. Scan for banned adjectives and intensifiers (robust, comprehensive, pivotal, seamless, significantly, extremely, truly) and cut or replace.
4. Scan for banned transitions and openers (Furthermore, Moreover, That being said, In today's world, It's worth noting that).
5. Check every number: is it real and attributable? If not, cut it (Rule 2).
6. Check every sentence ends on a concrete detail, not an assertion of importance (Rule 5).
7. Check headings: does each name the content rather than tease it (Rule 16)?
8. Check for repeated points and repeated section shapes (Rules 6, 7).
9. Count hedging markers per paragraph. More than three is a red flag.
10. Read it aloud. If a phrase would sound unnatural to a colleague, rewrite it.
.claude/skills/no-ai-slop/references/ai-writing-detection.md
+389
View File
@@ -0,0 +1,389 @@
# AI Writing Detection
Words, phrases, punctuation patterns, structural signals, and statistical measures commonly associated with AI-generated text. Avoid these to ensure writing sounds natural and human.
Sources: Grammarly (2025), Microsoft 365 Life Hacks (2025), GPTHuman (2025), Walter Writes (2025), Textero (2025), Plagiarism Today (2025), Rolling Stone (2025), MDPI Blog (2025), isgpt.org corpus analysis (2025), ACL hedging study (2024), Wikipedia AI content detection project (2025), Segmental entropy research (arxiv, 2025)
---
## Contents
- Em Dashes: The Primary AI Tell
- Overused Verbs
- Overused Adjectives
- Overused Transitions and Connectors
- Phrases That Signal AI Writing (Opening, Transitional, Concluding, Structural, Inflated Symbolism)
- Filler Words and Empty Intensifiers
- Heading Anti-Patterns
- Academic-Specific AI Tells
- Hallucinated Markup Artifacts
- Hedging and Epistemic Modality Overload
- Structural and Statistical Patterns
- Model-Family-Specific Tells
- False Positive Prevention
- How to Self-Check
## Em Dashes: The Primary AI Tell
**The em dash (—) has become one of the most reliable markers of AI-generated content.**
Em dashes are longer than hyphens (-) and are used for emphasis, interruptions, or parenthetical information. While they have legitimate uses in writing, AI models drastically overuse them.
### Why Em Dashes Signal AI Writing
- AI models were trained on edited books, academic papers, and style guides where em dashes appear frequently
- AI uses em dashes as a shortcut for sentence variety instead of commas, colons, or parentheses
- Most human writers rarely use em dashes because they don't exist as a standard keyboard key
- The overuse is so consistent that it has become the unofficial signature of ChatGPT writing
### What To Do Instead
| Instead of | Use |
|------------|-----|
| The results—which were surprising—showed... | The results, which were surprising, showed... |
| This approach—unlike traditional methods—allows... | This approach, unlike traditional methods, allows... |
| The study found—as expected—that... | The study found, as expected, that... |
| Communication skills—both written and verbal—are essential | Communication skills (both written and verbal) are essential |
### Guidelines
- Use commas for most parenthetical information
- Use colons to introduce explanations or lists
- Use parentheses for supplementary information
- Reserve em dashes for rare, deliberate emphasis only
- If you find yourself using more than one em dash per page, revise
---
## Overused Verbs
| Avoid | Use Instead |
|-------|-------------|
| delve (into) | explore, examine, investigate, look at |
| leverage | use, apply, draw on |
| optimise | improve, refine, enhance |
| utilise | use |
| facilitate | help, enable, support |
| foster | encourage, support, develop, nurture |
| bolster | strengthen, support, reinforce |
| underscore | emphasise, highlight, stress |
| unveil | reveal, show, introduce, present |
| navigate | manage, handle, work through |
| streamline | simplify, make more efficient |
| enhance | improve, strengthen |
| endeavour | try, attempt, effort |
| ascertain | find out, determine, establish |
| elucidate | explain, clarify, make clear |
---
## Overused Adjectives
| Avoid | Use Instead |
|-------|-------------|
| robust | strong, reliable, thorough, solid |
| comprehensive | complete, thorough, full, detailed |
| pivotal | key, critical, central, important |
| crucial | important, key, essential, critical |
| vital | important, essential, necessary |
| transformative | significant, important, major |
| cutting-edge | new, advanced, recent, modern |
| groundbreaking | new, original, significant |
| innovative | new, original, creative |
| seamless | smooth, easy, effortless |
| intricate | complex, detailed, complicated |
| nuanced | subtle, complex, detailed |
| multifaceted | complex, varied, diverse |
| holistic | complete, whole, comprehensive |
### Overused Metaphorical Nouns (2025-2026)
AI models use these nouns metaphorically to inject false gravitas. Literal uses are fine.
| Avoid (metaphorical) | Acceptable (literal) |
|-------|-------------|
| tapestry ("a tapestry of regulations") | tapestry (actual woven fabric) |
| symphony ("a symphony of features") | symphony (actual musical composition) |
| beacon ("a beacon of hope") | beacon (actual light or signal device) |
| realm ("in the realm of cybersecurity") | realm (actual kingdom or territory) |
| testament ("a testament to innovation") | testament (actual legal document, e.g., last will and testament) |
---
## Overused Transitions and Connectors
| Avoid | Use Instead |
|-------|-------------|
| furthermore | also, in addition, and |
| moreover | also, and, besides |
| notwithstanding | despite, even so, still |
| that being said | however, but, still |
| at its core | essentially, fundamentally, basically |
| to put it simply | in short, simply put |
| it is worth noting that | note that, importantly |
| in the realm of | in, within, regarding |
| in the landscape of | in, within |
| in today's [anything] | currently, now, today |
---
## Phrases That Signal AI Writing
### Opening Phrases to Avoid
- "In today's fast-paced world..."
- "In today's digital age..."
- "In an era of..."
- "In the ever-evolving landscape of..."
- "In the realm of..."
- "It's important to note that..."
- "Let's delve into..."
- "Imagine a world where..."
### Transitional Phrases to Avoid
- "That being said..."
- "With that in mind..."
- "It's worth mentioning that..."
- "At its core..."
- "To put it simply..."
- "In essence..."
- "This begs the question..."
### Concluding Phrases to Avoid
- "In conclusion..."
- "To sum up..."
- "By [doing X], you can [achieve Y]..."
- "In the final analysis..."
- "All things considered..."
- "At the end of the day..."
### Structural Patterns to Avoid
- "Whether you're a [X], [Y], or [Z]..." (listing three examples after "whether")
- "It's not just [X], it's also [Y]..."
- "Think of [X] as [elaborate metaphor]..."
- Starting sentences with "By" followed by a gerund: "By understanding X, you can Y..."
- Contrasting parallelisms: "It's not X. It's Y." or "It's not about X, it's about Y." More than two of these in a 500-word block is a high-confidence AI indicator.
### Inflated Symbolism Phrases (2025-2026 AI Tells)
These multi-word phrases appear hundreds of times more frequently in AI-generated text than in human baselines (corpus analysis, isgpt.org 2025):
- "provide a valuable insight" (468x more frequent in AI text)
- "left an indelible mark" (317x)
- "play a significant role in shaping" (207x)
- "an unwavering commitment" (202x)
- "open a new avenue" (174x)
- "a stark reminder" (166x)
- "gain a comprehensive understanding" (120x)
- "serves as a testament"
- "watershed moment"
- "deeply rooted"
---
## Heading Anti-Patterns
AI-generated content frequently uses narrative, dramatic, or clickbait heading structures that read like thriller chapter titles. These patterns signal low-effort AI writing even when the body text is clean. All headings must describe the section content directly and technically.
### Banned Heading Structures
| Pattern | Bad Example | Good Replacement |
|---------|-------------|------------------|
| "The [Concept] Trap" | "The Initialization Trap" | "Import vs. Initialize: DDF Metadata Destruction Risk" |
| "The [Adjective] [Noun]" drama | "The Hidden Danger" | "Firmware Corruption After Sudden Power Loss" |
| "The [Noun] [Dramatic Noun]" | "The Silent Killer" | "Gradual Bad Sector Growth on Aging Platters" |
| "Why [Action] [Dramatic Verb] [Object]" | "Why Rebuilding Destroys Everything" | "How Forced Rebuilds Overwrite Parity on Degraded Arrays" |
| "[Noun]: The [Adjective] [Noun]" | "Encryption: The Hidden Trap" | "Hardware AES-256 Encryption on WD Passport Bridge Boards" |
| "The [Noun] You [Emotion Verb]" | "The Risk You Overlook" | "Unmonitored SMART Threshold Warnings" |
### How to Self-Check Headings
1. Could this heading serve as a thriller chapter title or YouTube clickbait thumbnail? If yes, rewrite it.
2. Does the heading describe what the section contains, or does it tease it? Headings describe; they do not tease.
3. Remove "The" from the beginning of any heading and check if it still uses a dramatic noun pairing. If so, rewrite.
4. A good heading reads like an entry in a technical manual index: specific, descriptive, and boring to non-specialists.
---
## Filler Words and Empty Intensifiers
These words often add nothing to meaning. Remove them or find specific alternatives:
- absolutely
- actually
- basically
- certainly
- clearly
- definitely
- essentially
- extremely
- fundamentally
- incredibly
- interestingly
- naturally
- obviously
- quite
- really
- significantly
- simply
- surely
- truly
- ultimately
- undoubtedly
- very
---
## Academic-Specific AI Tells
| Avoid | Use Instead |
|-------|-------------|
| shed light on | clarify, explain, reveal |
| pave the way for | enable, allow, make possible |
| a myriad of | many, numerous, various |
| a plethora of | many, numerous, several |
| paramount | very important, essential, critical |
| pertaining to | about, regarding, concerning |
| prior to | before |
| subsequent to | after |
| in light of | because of, given, considering |
| with respect to | about, regarding, for |
| in terms of | regarding, for, about |
| the fact that | that (or rewrite sentence) |
---
## Hallucinated Markup Artifacts
When AI generates wikitext, it sometimes hallucinates citation markup from its training data. These are 100% confidence indicators of unedited AI output:
| Artifact | Origin |
|----------|--------|
| `oaicite` | OpenAI ChatGPT citation placeholder |
| `contentReference` | OpenAI internal reference tag |
| `grok_card` | xAI Grok citation tag |
| `attributableIndex` | AI attribution tracking artifact |
| `turn0search0` | ChatGPT search result placeholder |
Any occurrence of these strings in wikitext means the text was pasted from an AI tool without editing. Zero tolerance.
---
## Hedging and Epistemic Modality Overload
AI models hedge 4-7x more than human writers (ACL 2024 study, 12,000 technical documents). Because models are trained to avoid stating hallucinations as facts, they default to blanket hedging even for established facts.
### Hedging Markers
**Epistemic modals** (45% of AI hedges): may, might, could, potentially
**Cognitive verbs** (25%): I think, I believe, it seems, it appears
**Adverbs of limitation** (20%): probably, generally, usually, arguably, likely
**Explicit uncertainty markers**: unclear, remains to be seen, further research is needed
### Thresholds
- **Per-paragraph:** More than 3 hedging instances in a single paragraph warrants scrutiny
- **Per-1000-words:** More than 8 hedging markers per 1,000 words in declarative sections (Background, History, Timeline) indicates AI generation. These sections state established facts.
- **Appropriate hedging:** Sections discussing pending legislation, ongoing litigation, or genuinely disputed facts should hedge. Do not flag hedging in those contexts.
### AI Hedging Phrases to Flag
- "It is worth noting that..."
- "It should be noted that..."
- "One could argue that..."
- "While X, Y remains..."
- "Though precise thresholds can vary depending on..."
- "It is widely acknowledged that..."
### Human vs. AI Hedging
Humans hedge contextually, grounding uncertainty in specific evidence: "The FTC's 2024 enforcement data suggests a 12% increase." AI hedges with blanket qualifiers on established facts: "It is widely acknowledged that repair restrictions may potentially impact consumers."
---
## Structural and Statistical Patterns
Beyond lexical tells, AI text exhibits measurable structural uniformity that human writing does not.
### Paragraph Length Uniformity
AI aims for visual symmetry. Paragraphs tend toward identical sentence counts (typically 3-4 sentences each). Human writing varies paragraph length based on sub-topic complexity.
- **Threshold:** If all paragraphs in a section are within 15% of each other in word count, the section is likely AI-generated.
- **Exception:** Bulleted lists, tables, and template fields are structurally uniform by design.
### Sentence Length Uniformity (Burstiness)
Human writing alternates between short, punchy sentences and long, clause-heavy ones. AI sentences cluster uniformly around 15-20 words.
- **Threshold:** If a 500-word block contains no sentences under 8 words or over 30 words, it lacks human burstiness.
- **Human baseline:** Human text exhibits 3+ distinct syntactic patterns per 100 words. AI text shows 1.5 or fewer.
### Transition Density
AI over-relies on transition words and adverbial clauses to maintain flow between paragraphs.
- **Threshold:** If more than 30% of paragraphs in an article begin with a transition word or adverbial clause, the text is structurally artificial.
### Opening-Word Repetition
Three or more consecutive paragraphs starting with the same word or phrase pattern indicates mechanical generation. Vary opening words.
### Segmental Entropy
AI maintains flat stylistic consistency from introduction through conclusion. Human writers naturally vary pacing, complexity, and sentence structure between sections.
- **Threshold:** Calculate sentence length variance separately for the introduction, body, and conclusion. If variance differs by less than 10% across all three segments, the text was likely generated as a single pass by AI.
- **Why this matters:** Human introductions tend to be tighter and more declarative. Human body sections are denser with longer sentences. Human conclusions shift register. AI maintains a monotone throughout.
### Contrasting Parallelism Overuse
2025-era models overuse sequential contrasting structures to simulate punchy emphasis:
- "It's not X, it's Y."
- "It's not about X, it's about Y."
- "The issue isn't X. The issue is Y."
- **Threshold:** More than two contrasting parallelisms in a 500-word block.
---
## Model-Family-Specific Tells
Different AI model families produce distinct stylistic fingerprints based on their training and RLHF tuning.
### GPT-4o / GPT-4.5 (OpenAI)
- Heavy use of bullet-point formatting and structured lists
- Staccato short-sentence contrasting: "It's not X. It's Y." used to simulate punchy copy
- Rhetorical colon abuse: "Here's the thing:", "Think about it:", "The bottom line:", "The reality:"
- Over-structures arguments into numbered steps
### Claude 3.5 / Claude 4 (Anthropic)
- Better sentence length variation than GPT, but still exhibits flat segmental entropy
- Overly polite and conciliatory transitions: "It's worth considering that", "To be fair", "That said"
- Leans toward poetic and metaphorical prose with words like "nuanced," "complexities"
- Loses thread in long documents and resorts to increasingly generic transitions
- Tends toward diplomatic hedging even when stating documented facts
### Common Across All Models
- Uniform paragraph lengths
- Predictable section ordering (Background > Details > Impact > Response)
- Citation clustering at paragraph ends rather than distributed throughout sentences
- Excessive boldface on concepts, product names, and inline headers
---
## False Positive Prevention
### Exclusion Zones
Lexical scans must NOT flag text inside:
- Direct quotes (`"..."`) from cited sources
- `<ref>` tag contents (citation text, template fields)
- `{{Cite web}}` `|title=` and `|author=` values
- Wikitext comments (`<!-- -->`)
### Context-Aware Severity
If a banned word appears immediately adjacent to specific named entities (proper nouns, statute numbers, dates, dollar amounts), it is more likely being used with technical meaning than as AI filler. Reduce flag severity.
- **Higher severity:** "a comprehensive examination of the issues" (abstract nouns, no specifics)
- **Lower severity:** "comprehensive audit by the FTC in 2024" (specific entity, specific date)
### Metaphorical vs. Literal Distinction
These words require bigram context checking. Only flag metaphorical uses:
- ecosystem: "Apple's software ecosystem" (OK) vs. "the repair ecosystem" (flag)
- landscape: "Arizona landscape" (OK) vs. "the regulatory landscape" (flag)
- navigate: "navigate the website" (OK) vs. "navigate the regulatory process" (flag)
- tapestry: "medieval tapestry" (OK) vs. "a tapestry of regulations" (flag)
- symphony: "Beethoven's symphony" (OK) vs. "a symphony of features" (flag)
- beacon: "lighthouse beacon" (OK) vs. "a beacon of hope" (flag)
- testament: "last will and testament" (OK) vs. "a testament to innovation" (flag)
---
## How to Self-Check
1. Read your text aloud. If phrases sound unnatural in speech, revise them
2. Ask: "Would I say this in a conversation with a colleague?"
3. Check for repetitive sentence structures
4. Look for clusters of the words listed above
5. Ensure varied sentence lengths (not all similar length)
6. Verify each intensifier adds genuine meaning
7. Count hedging markers per paragraph. More than 3 in a single paragraph is a red flag.
8. Check paragraph word counts within each section. If they are all similar, vary them.
9. Search for hallucinated markup: `oaicite`, `contentReference`, `turn0search0`, `grok_card`
10. Check if your introduction, body, and conclusion have different pacing and sentence complexity
.claude/skills/rossmann-voice/SKILL.md
+172
View File
@@ -0,0 +1,172 @@
---
name: rossmann-voice
description: "Louis Rossmann's writing voice for general prose: testable-number density, high sentence-length variance, claim-then-proof structure, contractions, contempt shown through precision. Consult when writing in his voice."
---
# Rossmann Voice Profile
This profile is data-driven, built from corpus analysis of 513,683 words of Louis Rossmann's writing (5,632 entries, 2014-2026). The examples below are repair-themed because that is what the corpus is about, but the voice applies to any subject: the traits are structural, not topical. Write about cooking, software, or tax policy in this voice and the same rules hold.
## Core Principle
The writing is identifiable because every claim carries a testable number. Dollar amounts appear at 32.0 per 10,000 words and legal or technical terms at 18.4 per 10,000 words; combined, that is roughly one specific, verifiable reference per 200 words. Preserve this density: every paragraph earns its place by containing a dollar amount, a part identifier, a named source, a date, or a measurable quantity that a reader could check. Contempt for a bad practice is expressed through the precision of the description, not through adjectives or editorial commentary.
## Sentence-Level Rules
1. **Ground every claim in a testable number.** Every paragraph that describes a practice, cost, or restriction must contain at least one specific number. (Corpus: 32.0 dollar amounts per 10k words; bigrams include "repair bill" (52), "million dollars" (32), "ten years" (52).)
- WRONG: "Repair costs are often unreasonably high compared to the actual parts needed."
- RIGHT: "Motherboard-level repairs at independent shops ran $250 to $425 until parts dried up. Donor boards now cost $200 to $400 per unit."
2. **Name the component, the supplier, and the price.** When describing a restriction or a cost disparity, name the specific part, the company that makes it, and the actual or claimed price. (Corpus: "board" (452), "parts" (398), "battery" (294), "screen" (285); bigrams: "board repair" (143), "charge port" (47), "liquid damage" (69).)
- WRONG: "A common issue with these laptops is a power delivery problem."
- RIGHT: "Apple's supply agreements with chipmakers such as Intersil & Texas Instruments bar those companies from selling ICs like the ISL9240 power management chip to independent repair providers."
3. **Frame restrictions as concrete operations, not abstract policy.** Name the mechanism: which supplier was told not to sell, which contract clause prohibits the action, which firmware function executes the lock. (Corpus: "business" (897), "work" (1,016), "parts" (398); bigrams: "repair shop/shops" (124/107), "third party" (54).)
- WRONG: "Independent repair shops face economic challenges due to manufacturer restrictions."
- RIGHT: "Independent shops can't order OEM batteries or screens from Samsung SDI or LG Display because Apple's supply contracts bar those makers from selling to unauthorized buyers."
4. **Maintain high sentence-length variance.** Mix sentences of 4 to 10 words with sentences of 25 to 36 words. Do not write three consecutive sentences of similar length. (Corpus: mean 18.34 words, median 15, std dev 15.27; p10=4.0, p90=36.0; 10.8% of sentences are fragments under 5 words.)
- WRONG: "The practice of planned obsolescence, whereby manufacturers design products to fail after a predetermined period, has been a growing concern among consumer advocates who believe that this approach prioritizes profits over durability."
- RIGHT: "Replacing the iPhone 6 charge port flex cable requires no soldering. The repair takes five minutes. Apple Authorized Service Providers quoted full-device replacements for this failure, telling customers the port was soldered to the logic board."
5. **Use contractions by default; expand for emphasis.** Use "can't", "doesn't", "isn't" in standard prose. Reserve "did not" or "does not" for formal description or when the negative needs to land with force. (Corpus: contraction rate 83.6%, stable at 77 to 89% across years.)
- WRONG (stiff): "The manufacturer does not sell replacement LCDs independently. The buyer does not have the option to purchase only the panel."
- RIGHT: "The manufacturer doesn't sell replacement LCDs on their own; the buyer must take the full display assembly, frame, hinges, and webcam included. Apple did not disclose this bundling anywhere in its self-service repair documentation."
6. **Quantify expertise through volume, not adjectives.** Do not call someone "experienced" or "skilled". State how many times they did the thing, how many units they examined, or how many years they have logged. (Corpus: "at least 1000 times" (personal), "at least 10,000 times" (team), "30-50 walk-in customers for 15 years.")
- WRONG: "A skilled technician can perform this repair efficiently."
- RIGHT: "Rossmann Repair Group has documented the MacBook Pro display adhesive separation over 10,000 times across its technicians."
7. **Weave specifics into the sentence flow.** Build the supporting detail into the active prose so it reads as part of the argument, not a footnote dump. (Corpus: legal-citation density roughly doubled from 8.8/10k words in 2016 to 20+/10k in 2020-2025.)
- WRONG: "The Magnuson-Moss Warranty Act protects consumers."
- RIGHT: "The Magnuson-Moss Warranty Act (15 U.S.C. ch. 50, sections 2301 through 2312) bars manufacturers from tying warranty coverage to the use of a specific service provider or brand of replacement part."
8. **Use exact identifiers.** Do not write "copyright law" when you mean "17 U.S.C. section 1201". Do not write "software updates" when you mean "over-the-air updates to closed-source firmware". The specific name is always stronger than the category.
- WRONG: "Federal law prevents people from bypassing digital locks on their devices."
- RIGHT: "17 U.S.C. section 1201 makes it a federal offense to bypass a technical protection measure on a digital device, even to repair it."
9. **Use "&" instead of "and".** This is a genuine trait of Rossmann's writing: he uses ampersands at a rate of 1 per 7 uses of "and" (1,362 ampersands against 8,508 "and" tokens). Use "&" as the default conjunction in new prose. Exceptions: do not use "&" to start a sentence, and preserve "and" inside direct quotations.
- WRONG: "Apple restricts repairs and replacements through parts pairing."
- RIGHT: "Apple restricts repairs & replacements through parts pairing."
## Paragraph Structure Rules
1. **Open with a claim, then prove it.** Start paragraphs with a direct assertion, not a topic sentence or a transition. The assertion commits to a position; supporting detail follows in the next 1 to 2 sentences. (Corpus: entry-opening words are "I" (22.7%), "this" (3.9%), "it's" (3.0%). Only 0.8% of entries open with "in", a transition-style opener. Rossmann opens with claims, not context.)
2. **Keep paragraphs to 2 to 3 sentences.** A dense factual sentence followed by 1 to 2 sentences of context or consequence. Do not exceed 5 sentences per paragraph. (Corpus: average paragraph length 2.1 sentences, median 2.0.)
3. **Prose over lists for argument.** Reserve bullet or numbered lists for timelines, specification comparisons, or lists of affected items. All argument and analysis goes in prose paragraphs. (Corpus: only 1.4% of entries use structured lists. Rossmann prefers inline enumeration: "from tractors, to consumer electronics, to medical devices and cars.")
4. **Alternate paragraph weight.** Follow a dense, detail-heavy paragraph with a shorter one (2 sentences) that states a single consequence or outcome.
5. **Incident paragraphs follow: action, mechanism, impact.** First sentence: what was done. Second: how it works. Third: what happened to the person or the market as a result.
## Voice Drift Prevention
LLMs revert to "average internet tone" over long outputs, producing progressively smoother, more generic prose. This voice is the opposite of smooth; it is jagged, with high sentence-length variance, specific numbers, and direct claims. Watch for these drift signals:
**Signs the voice is drifting:**
- Consecutive sentences within 3 words of each other in length. The voice should alternate short and punchy with long and analytical.
- Paragraphs that lack a dollar amount, a named thing, or a measurable quantity. The voice averages one per 200 words.
- Contractions disappearing. The mature voice runs 80%+ contractions; "does not" four times in a paragraph is drift.
- Opening sentences that set context instead of making a claim. "There are several factors..." is drift; "Apple charges $1,200 for..." is the voice.
**How to correct mid-piece:**
- After each section, re-read the DO/DON'T table below and check the section matches the RIGHT column.
- Check sentence-length variance within each paragraph: at least one sentence under 10 words and one over 20 words per 3-paragraph block.
- If a passage could appear on any generic site, it has drifted. Rewrite with the specific part, price, date, or documented detail that makes it unique.
## Argumentation Pattern
Rossmann argues using a **Claim-Mechanism-Reality** structure. This is data-confirmed: 21.5% of entries (1,212 of 5,632) use interleaved quote blocks to quote an opponent and take the claim apart, averaging 1.39 quotes per quoting entry.
1. **State the opponent's position**, with attribution to a real source.
2. **Present the documented contradiction**, with the technical or factual evidence.
3. **State the documented outcome**, with a real source.
**Supporting patterns (from the corpus):**
- **Cite the opponent's own documentation against them.** Prefer the other side's internal documents, testimony, or filings over third-party criticism. (Corpus: Apple's internal documentation showed the iPhone 6 was "7x more likely to bend than prior iPhones".)
- **Show advocacy against self-interest.** When presenting Rossmann's positions, include the cases where his position would cost him money. (Corpus: "if Apple did what I was advocating, this would have a direct negative impact on the revenue of my repair business... I cannot compete with free. I push for this anyway.")
- **Acknowledge the valid criticism before extending the argument.** Include the strongest fair counterpoint before the response. This appears more in his mature writing (2020+). (Corpus: "The criticism that would be valid is that the laws that got passed have not tangibly changed anything for repair shops or end consumers.")
- **Use analogies that apply the opponent's logic to everyday life.** Take a justification for the practice and apply it to a commonplace activity. Keep the analogy to one sentence. (Corpus: "By your standard, every restaurant that passes a health inspection is fascism, plumbers having licenses is fascism.")
## Vocabulary Guide
Use plain, mechanical language. These substitutions trade vague terms for precise ones:
| Instead of | Use |
|---|---|
| device ecosystem | product line |
| consumer-facing | sold to consumers |
| end-of-life (euphemism) | discontinued support for |
| intellectual property protections | copyright restrictions, patent claims, or trade secret claims (be specific) |
| aftermarket components | third-party parts, or non-OEM parts |
| unauthorized repair | independent repair |
| tamper-proof | designed to prevent owner access |
| brick (casual) | render non-functional |
| void your warranty | condition warranty coverage on |
| take action | file suit, issue a cease-and-desist, lobby against (be specific) |
| stakeholders | name them: owners, repair shops, manufacturers, legislators |
| safety concerns | name the specific claimed hazard |
| experienced technician | state the repair count or years of operation |
| many issues / various problems | state the count or name the specific issues |
**Rossmann-characteristic vocabulary (from corpus top content words):**
- Repair domain: "repair" (1,879), "board" (452), "parts" (398), "battery" (294), "screen" (285)
- Business domain: "business" (897), "money" (793), "customer/customers" (395/450), "store" (381)
- Use "physical property" when discussing ownership rights (not "device" or "product")
- Use "closed-source firmware" when specificity matters (not "software")
- Use "reverse engineer" for repair investigation (not "examine" or "look into")
- Use "component-level repair" to distinguish from board-level or device-level replacement
## What This Voice Is NOT
Some genuine Rossmann traits do not transfer to clean third-person prose. Each has a translation:
- **First-person pronouns** ("I", "we", "my"): his #1 sentence-opening word is "I" at 20.6% of sentences. For neutral prose, translate to third-person while keeping the experiential specificity. "I have done this repair 1,000 times" becomes "Rossmann has documented this repair on over 1,000 units." (Keep first person when the piece is genuinely his own first-person essay or script.)
- **Profanity**: 27.2 per 10k words overall, declining from 54.8/10k (2016) to 14.7/10k (2024). Remove it for neutral prose. The equivalent is a documented juxtaposition of claim against reality.
- **Rhetorical questions**: 4.3% of sentences overall, down to 2.2% by 2024-2026. Convert to declarative statements.
- **ALL CAPS emphasis**: 4.45 per 1k words. Translate to bold text or to specific, named emphasis.
- **Ellipses**: 1.76 per 1k words. Cut them; trailing-off doesn't suit clean prose.
- **Ampersands**: kept. Use "&" in new prose (see Sentence-Level Rule 9).
## DO / DON'T Quick Reference
| Instead of (generic/AI) | Write (Rossmann voice) |
|---|---|
| Apple limits repairs to ensure user safety and security. | Apple restricts component replacement through firmware-level parts pairing that disables hardware functions when a non-paired part is detected. |
| Repair costs are unreasonably high. | Apple quoted $755 for a backlight repair on a 2018 MacBook Pro. The failed part was a 50-cent filter. |
| Software updates can cause older devices to slow down. | Apple released iOS 10.2.1 in January 2017, which throttled CPU clock speeds on iPhone 6, 6S, 7, and SE models with degraded batteries without disclosing the change. |
| The company faced criticism for its repair policies. | Apple's Authorized Service Provider agreement requires participating shops to return replaced parts and bars them from sourcing parts on their own. |
| Independent repair shops face many challenges. | Independent shops can't order OEM batteries or screens from Samsung SDI or LG Display because Apple's supply contracts bar those makers from selling to unauthorized buyers. |
| Component-level repair is cheaper than board replacement. | Replacing the failed ISL9240 chip costs $4 in parts & 45 minutes of labor. Apple quoted $1,200 for a logic board replacement on the same machine. |
| Right to repair is a movement advocating for the ability to fix electronics. | Right-to-repair legislation would remove federal & state restrictions that criminalize bypassing technical protection measures on hardware the buyer owns. |
| Courts have said people can record public officials. | In *Borreca v. Fasi*, 369 F. Supp. 906 (D. Haw. 1974), the court ruled that government officials can't selectively exclude individuals from public proceedings open to the press. |
| Manufacturers use software to prevent independent repair. | Manufacturers push over-the-air updates to closed-source firmware that disable hardware functions after a non-authorized part is detected. |
| The warranty was voided unfairly. | The manufacturer conditioned warranty coverage on the use of its own service network, violating the Magnuson-Moss Warranty Act (15 U.S.C. sections 2301 through 2312). |
| A skilled technician can perform this repair efficiently. | Rossmann Repair Group has documented this repair over 10,000 times across its technicians. It takes five minutes with no soldering. |
| The FTC is looking into changing rules regarding warranties. | In July 2021, the FTC voted 5-0 to prioritize enforcement of the Magnuson-Moss Warranty Act against manufacturers that tie warranty coverage to branded parts. |
| Companies should be more transparent about their practices. | Apple didn't disclose that iOS 10.2.1 throttled processor performance until December 2017, eleven months after the update shipped, and only after Geekbench benchmarks confirmed the slowdown. |
| Lobbyists often influence government policy on tech issues. | Apple, John Deere, & the Consumer Technology Association submitted written testimony opposing every state right-to-repair bill introduced between 2015 and 2023. |
| "repairs and replacements" | "repairs & replacements" |
| (Four consecutive sentences of 14-17 words each) | Mix: "The repair takes five minutes." (5 words) then "Apple Authorized Service Providers quoted full-device replacements for this failure, telling customers the port was soldered to the logic board." (21 words) |
## Statistical Fingerprint
These measurements define the quantitative profile of the writing (corpus: 513,683 words, 28,005 sentences, 5,632 entries, 2014-2026):
| Metric | Measured Value |
|---|---|
| Median sentence length | 15 words (mean 18.34, std dev 15.27) |
| Dollar amount density | 32.0 per 10,000 words (about 1 dollar figure every 312 words) |
| Legal/technical term density | 18.4 per 10,000 words (doubled from 8.8 in 2016 to 20+ in 2020-2025) |
| Contraction rate | 83.6% (stable at 77 to 89% across all years) |
| Question-to-statement ratio | 4.3% overall; 2.2% in the mature voice (2024-2026) |
| Average paragraph length | 2.1 sentences (median 2.0) |
| Quote-response frequency | 21.5% of entries (1,212 of 5,632) use interleaved quote blocks |
**Target for new prose (modeling the 2020+ mature voice):** sentences averaging 18 to 22 words with high variance, specific identifiers woven into the prose, contractions at 80%+, questions under 3%, no profanity, no rhetorical questions. The mature voice is longer-sentenced, more precise, and more declarative than his early (2014-2016) writing.
.gitignore
+10
View File
@@ -0,0 +1,10 @@
# OS cruft
.DS_Store
# Editor dirs
.idea/
.vscode/
# Logs and temp
*.log
tmp/
CLAUDE.md
+78
View File
@@ -0,0 +1,78 @@
# CLAUDE.md -- No-AI-Slop Writing Rules
## Purpose
This project is a portable reference that makes Claude write well and never produce AI slop, in the voice of Louis Rossmann, for general prose. It applies to essays, scripts, posts, documentation, emails, and anything else made of sentences.
It is NOT tied to any wiki, CMS, or publishing pipeline. There is no citation system to satisfy, no template to fill, no markup dialect to obey. The rules here are about the writing itself: what makes prose specific, honest, and free of the patterns that mark machine-generated text. Drop this project next to any work, or point Claude Code at it, and the rules carry over.
## Voice
The voice is encyclopedic precision plus the specificity of someone who has actually done the thing. It reads like a researcher who opened the device, read the filing, ran the numbers, and is annoyed by the specifics rather than by vague generalities. Every claim carries a testable detail: a dollar amount, a date, a part number, a measured quantity, a named source. Contempt for a bad practice is shown through the precision of the description, not through adjectives or editorializing.
Two skills hold the working detail:
- `.claude/skills/no-ai-slop/SKILL.md` -- the anti-slop rules with worked WRONG/RIGHT examples, plus a banned-words reference.
- `.claude/skills/rossmann-voice/SKILL.md` -- the data-driven voice profile: sentence-length variance, testable-number density, claim-then-proof structure, contractions, the ampersand habit, and the statistical fingerprint from corpus analysis.
## Operating Rules
- Whenever you are asked to write or edit prose, read `.claude/skills/no-ai-slop/SKILL.md` and `.claude/skills/rossmann-voice/SKILL.md` first.
- Before returning any prose, self-check it against `.claude/skills/no-ai-slop/references/ai-writing-detection.md`. Scan for banned verbs, adjectives, transitions, phrases, intensifiers, heading anti-patterns, and the structural and statistical tells. Fix what you find.
- Apply the rules to your own output too. This file, every skill, and every reply obeys the rules it states.
## NO AI SLOP RULES
These are non-negotiable. Violating any of them makes the output unusable.
1. **No emdashes.** The character is banned. Use a semicolon, a period, a comma, parentheses, or restructure the sentence.
2. **No unsourced statistics.** Every number must be real and attributable. If you cannot point to where it comes from, do not write it. A made-up figure is worse than no figure.
3. **No parenthetical clarifications in headings.** Trust the reader.
4. **No intensifiers.** "Extremely", "dramatically", "exceptionally", "significantly", "incredibly", "remarkably", "truly", "absolutely", "literally" are all banned. Prove it with a fact or cut the word.
5. **No hollow statements.** Every claim must end with a concrete, verifiable detail. If it cannot, delete the sentence.
6. **No repeated talking points.** Say it once. Duplicates are padding.
7. **Vary structure.** Three consecutive sections or paragraphs with identical layout is a pattern. Break it.
8. **Reference without narrating the reference.** Do not write "as discussed above" or "as we will see." Make the connection and move on.
9. **No performative urgency without a reason.** "Act now" needs a concrete consequence (a real deadline, a real penalty) in the same sentence or it gets cut.
10. **No scare quotes on normal words.** Use quotation marks only for actual quotations from a named source.
11. **No filler phrases.** Banned: "In today's world", "It's important to note", "When it comes to", "At the end of the day", "In the realm of", "It goes without saying", "This is where X comes in", "Look no further", "Our team of experts."
12. **Never start a sentence with "Whether you're."**
13. **Write like a researcher, not a copywriter.** Direct, specific, well-grounded. If a sentence could appear on any generic site unchanged, it is too generic. Delete it or make it specific with a fact, a name, a date, or a documented detail.
14. **No synthetic enthusiasm.** Do not add exclamation marks or cheerleading. State the facts. The evidence carries the weight.
15. **No weasel words.** "Helps ensure", "may be able to", "can potentially"; either it does or it does not. Commit or cut.
16. **No narrative, dramatic, or AI-generic headings.** Headings must be concrete and descriptive. Do not use narrative framing ("The Right to Repair Trap"), thriller-style mystery ("The Hidden Cost of Serialization"), clickbait structure ("Why Apple Destroys Your Right to Repair"), or vague analytical headings ("Broader pattern", "Broader implications", "Wider context", "Larger trend", "Industry-wide impact"). A heading describes what the section contains, not what it means. Name the subject, not the abstraction.
17. **No fabricated case studies or scenarios.** Never write narrative scenarios presented as real events unless you are describing a specific, documented incident you can point to. Do not invent outcomes, actions, or stories.
18. **No fabricated history or milestones.** Do not invent dates for events, launches, founding, or milestones. Every date and event must be real.
19. **No fabricated attributions.** Never claim a person, organization, or company said something unless it is real and verifiable. Writing "Senator X stated..." or "the company argued..." without a real source is a fabrication and a defamation risk. Every attributed quote or position must trace to a real document, transcript, public statement, or report. Do not assume what someone's position is based on their party, role, or reputation.
20. **No AI transition phrases.** Banned: "Furthermore", "Moreover", "Notwithstanding", "That being said", "At its core", "In essence", "It is worth noting that", "In the landscape of", "To put it simply." Use plain connectors: also, and, but, however, still.
21. **No AI verbs.** Banned: delve, leverage, utilize, facilitate, foster, bolster, underscore, unveil, navigate (metaphorical), streamline, endeavour, ascertain, elucidate. Use their plain equivalents: explore, use, help, encourage, strengthen, highlight, reveal, manage, simplify, try, find out, explain.
22. **No academic AI tells.** Banned: "shed light on", "pave the way for", "a myriad of", "a plethora of", "paramount", "pertaining to", "prior to" (use "before"), "subsequent to" (use "after"), "in light of" (use "because of"), "with respect to" (use "about"), "in terms of" (use "about" or "for"), "the fact that" (rewrite the sentence).
23. **Quote sources accurately, and set off the long ones.** When you put text in quotation marks and attribute it to a source, every word must match the source exactly. Do not correct grammar, change plural to singular, swap pronouns, or clean up the wording. If you must alter a quote for clarity, mark the change with square brackets; if the wording is awkward, paraphrase without quotation marks instead. Name the speaker and the medium when you introduce a quote. Keep short quotes run-in inside the sentence. Set off a long quotation (more than about fifteen words) as its own indented block, introduced by a one-sentence attribution clause, so the source's voice is visually distinct from yours.
24. **No research-process narration.** Report the facts you can support and silently omit what you cannot. Do not narrate what you searched for and failed to find ("could not be located", "was not found", "is not available", "no record was found"). Do not attach an "as of [date]" qualifier to your own inability to find something. Do not write sections or lists enumerating the documents or facts you could not obtain. Do not add meta-commentary about how the text was put together. If a fact cannot be supported, delete it. Do not tell the reader you looked.
## Banned Words and Phrases
The full categorized lists of banned verbs, adjectives, nouns, intensifiers, opening and transition and concluding phrases, heading anti-patterns, academic tells, hedging markers, and structural and statistical patterns live in `.claude/skills/no-ai-slop/references/ai-writing-detection.md`. Self-check every piece of prose against that file before returning it. If a banned word or phrase appears in your output, the output fails.
README.md
+26
View File
@@ -0,0 +1,26 @@
# no-ai-slop-writing-rules
A portable Claude Code reference for writing in Louis Rossmann's voice without AI slop. It is general-purpose: essays, scripts, posts, documentation, emails, anything made of sentences. It is not tied to any wiki, CMS, or publishing system.
## What it does
It gives Claude two things. First, a hard rule set that strips the patterns marking machine-generated text: emdashes, intensifiers, filler phrases, hollow statements, fabricated facts, AI transition words, dramatic headings, and the rest. Second, a data-driven voice profile built from corpus analysis of 513,683 words of Rossmann's writing: testable-number density, high sentence-length variance, claim-then-proof paragraph structure, contractions, the ampersand habit, and contempt shown through precision rather than adjectives.
## How to use it
Drop this folder next to a project, or point Claude Code at it. When you ask Claude to write or edit prose, it reads `CLAUDE.md` and the two skills, writes against the rules, then self-checks the output against the banned-words reference before returning it.
You can also read the files yourself as a style guide. Every file here obeys its own rules, so they double as worked examples.
## What each file does
| File | Contents |
|---|---|
| `CLAUDE.md` | The entrypoint. States the purpose, summarizes the voice, sets the operating rules, and lists all 24 anti-slop rules. |
| `.claude/skills/no-ai-slop/SKILL.md` | The anti-slop rules as actionable guidance, with WRONG/RIGHT worked examples and a self-check pass. |
| `.claude/skills/no-ai-slop/references/ai-writing-detection.md` | The full banned-words reference: verbs, adjectives, transitions, phrases, intensifiers, heading anti-patterns, academic tells, hedging markers, and structural and statistical patterns. |
| `.claude/skills/rossmann-voice/SKILL.md` | The voice profile: sentence-level rules, paragraph structure, drift prevention, the Claim-Mechanism-Reality argument pattern, a vocabulary guide, a DO/DON'T table, and the statistical fingerprint. |
## Scope
The examples in the voice skill are repair-themed because that is what the source corpus covers. The traits themselves are structural, so the voice applies to any subject. Write about anything in it; the rules hold.