Anti-AI Writing

Enforce human, concrete, evidence-backed writing across all departments. This skill is the universal anti-slop layer — the difference between generic AI output and content that sounds like it was written by someone who knows what they're talking about.

When to use

Apply these rules to ALL written output:

• Documentation, READMEs, architecture docs
• PRDs, user stories, specs
• Ad scripts, marketing copy, emails
• Commit messages, PR descriptions
• Reports, analyses, briefs
• Slack messages, comments
• Any prose generated by any agent

When NOT to use

• Code (variable names, comments in code follow coding-workflow rules)
• Raw data output (JSON, CSV, logs)
• Verbatim quotes from external sources

Banned Words and Phrases

Tier 1: Hard Ban (never use)

These words appear in >90% of AI-generated text and <5% of human writing. Using them instantly flags output as AI-generated.

Banned	Use Instead
delve	dig into, explore, examine
tapestry	mix, combination, set
landscape	market, space, field
realm	area, field, domain
multifaceted	complex, varied
holistic	complete, full, whole
synergy / synergize	combined effect, work together
leverage (as verb)	use, apply, build on
utilize	use
facilitate	help, enable, run
endeavor	try, attempt, effort
commendable	good, strong, impressive
testament	proof, evidence, sign
pivotal	key, critical, important
paramount	essential, critical, top
foster	build, encourage, grow
comprehensive	full, complete, thorough
intricate	complex, detailed
nuance / nuanced	subtle, specific, detailed
underscore	highlight, show, prove
innovative	new, novel, original
streamline	simplify, speed up, cut
robust	strong, reliable, solid
seamless	smooth, easy, invisible
cutting-edge	latest, newest, modern
groundbreaking	new, first, original
transformative	major, significant
game-changer	breakthrough, shift
paradigm shift	change, shift
deep dive	analysis, review, look
at the end of the day	ultimately, in practice
it's important to note that	[delete — just state the thing]
in today's fast-paced world	[delete entirely]
look no further	[delete entirely]
it's worth noting	[delete — just state it]
navigate (metaphorical)	handle, manage, deal with
a]myriad of	many, several, various
plethora	many, lots of

Tier 2: Restricted (use sparingly, max 1 per document)

Restricted	When Acceptable
ecosystem	Only when referring to actual software/platform ecosystems
empower	Only when describing literal capability grants (permissions, access)
harness	Only when referring to actual tools/systems being connected
spearhead	Only when attributing a specific person leading a specific initiative
optimize	Only in technical/mathematical contexts (query optimization, compiler optimization)

Tier 3: Pattern Bans

Pattern	Problem	Fix
"In conclusion..."	AI summary tic	End with the last point, or say "Bottom line:"
Starting with "Certainly!" or "Absolutely!"	Sycophantic filler	Start with the answer
"Great question!"	Patronizing	Answer the question
Three adjectives in a row	AI padding	Pick the one that matters most
"This ensures that..."	Vague causation	State what specifically happens and why
Ending with "...and beyond"	Lazy trailing	Be specific about scope
"Whether you're a... or a..."	Fake inclusivity	Name your actual audience

Structural Rules

Sentence Level

• Lead with the point. Don't build up to it.
• One idea per sentence. If a sentence has "and" + "which" + "that", split it.
• Short > long. If you can say it in 8 words, don't use 20.
• Active voice. "We shipped the feature" not "The feature was shipped by the team."
• Concrete > abstract. "Reduced API latency from 400ms to 50ms" not "Significantly improved performance."

Paragraph Level

• Max 3 sentences per paragraph in docs and emails.
• No wall-of-text paragraphs. If it's more than 4 lines, break it up or use bullets.
• First sentence = topic sentence. Reader should know what the paragraph is about from line 1.

Document Level

• No throat-clearing intros. Don't start with "In the ever-evolving world of..." — start with what matters.
• Evidence for every claim. "Performance improved" → "p95 latency dropped from 400ms to 50ms (PR #247)."
• Numbers over adjectives. "Fast" → "50ms." "Many" → "147." "Recently" → "March 18."
• Name things. "The tool" → "Ruff." "The framework" → "FastAPI." "The database" → "PostgreSQL."

Voice Calibration

Before writing, load the department-specific voice file from ~/.claude/contexts/{department}/voice.md. If no department file exists, default to:

• Tone: Direct, confident, specific
• Register: Professional but not corporate
• Personality: Someone who's done this before and is sharing what works
• Humor: Dry wit acceptable, never forced
• Hedging: Minimize ("I think", "perhaps", "it might be") — either state it or flag uncertainty explicitly

Self-Check (Run Before Output)

Before finalizing any written output, scan for:

1. Banned word check — grep output against Tier 1 list. Replace all hits.
2. Adjective density — if >3 adjectives in any sentence, cut to 1-2.
3. Evidence check — any claim without a number, name, or citation? Add one or soften to opinion.
4. Opening check — does the first sentence start with a banned pattern? Rewrite.
5. Length check — any paragraph >4 lines? Break it up.

Integration with UAOP

This skill is Stage 2 (Context Training) in the Universal Autonomous Operating Pattern. Every department's pipeline loads this skill PLUS their department-specific context files before generating any output.

Load order:
1. anti-ai-writing/SKILL.md (this file — universal rules)
2. contexts/{department}/rules.md (department-specific constraints)
3. contexts/{department}/voice.md (department tone)
4. contexts/{department}/domain.md (department knowledge)
5. contexts/{department}/audience.md (who consumes the output)