SuperAPI AI Directory

avoid-ai-writing

Writing & research

View repository

Updated 10-15 20:30

About this skill

Avoid AI Writing — Audit & Rewrite

You are editing content to remove AI writing patterns ("AI-isms") that make text sound machine-generated.

When to Use

  • When asked to "humanize" or "clean up" AI-generated text.
  • Before publishing agent-generated READMEs, changelogs, or documentation.
  • When auditing content for statistical AI tells.

Core Modes

  • rewrite (default): Flag AI-isms and rewrite text to fix them. Returns Issues, Rewritten text, What changed, and a Second-pass audit.
  • detect: Flag AI-isms only. No rewriting. Useful for CI/CD gates or audits.
  • edit: Edit a file in-place with minimal, targeted changes using the Edit tool.

Voice Profiles (Optional)

Independent of audience context; sets the persona.

  • casual: Contractions, short sentences (≤14 words), low jargon.
  • professional: Active voice, concrete claims, low hedging.
  • technical: Plain copulatives (is/has), one idea per sentence, imperative mood.
  • warm: Direct address ("you"), strong verbs, medium cadence (15-20 words).
  • blunt: Short declaratives, no padding, no hedges.

Context Profiles

Adjusts rule strictness for specific audiences.

  • linkedin: Social-first; relaxed on formatting, strict on P0 credibility killers.
  • blog: Default; all rules full strength.
  • technical-blog: Technical terms (robust, leverage, etc.) exempt; strict prose rules.
  • investor-email: Extra strict on promotional language and significance inflation.
  • docs: Clarity over voice; relaxed on hedging and copula avoidance.
  • casual: Slack/DM style; P0 credibility killers only.

Instructions

  1. Audit: Identify AI-isms using the references/patterns.md catalog.
  2. Apply Profile: Adjust strictness based on context and tone based on voice.
  3. Execute Mode:
    • rewrite: Return audit, full rewrite, change summary, and second-pass audit.
    • detect: Return severity-grouped issues (P0/P1/P2) and assessment.
    • edit: Apply minimal Edit tool changes to target file; return edit report.
  4. Convergence: If --iterate 2 is passed, repeat until no patterns remain (max 2 passes).

Severity Tiers

  • P0 — Credibility killers: Cutoff disclaimers, chatbot artifacts, vague attributions.
  • P1 — Obvious AI smell: Word-list violations, template phrases, "Let's" openers.
  • P2 — Stylistic polish: Generic conclusions, uniform length, transition overuse.

Rationalizations

Rationalization Reality
"AI writing is grammatically perfect, so it's fine." Perfection is a tell; natural human writing has varied rhythm and personality.
"I can just ask the LLM to 'sound human'." One-shot "sound human" prompts are vibes-based; this skill uses a structured audit.
"This will flag non-native speakers." Patterns are signals, not proof. Use judgment and pair with context.

Red Flags

  • Sanding away all personality in pursuit of "clean" prose (creating new uniformity).
  • Rewriting quoted material or code blocks.
  • Ignoring the Second-pass audit (patterns often survive the first edit).

References

  • references/patterns.md — Exhaustive catalog of AI patterns and vocabulary.

See Also

  • ui-ux-optimize — For optimizing UI/UX specific copy and layouts.
  • readme-best-practices — For documentation structure and discovery.

Voice & Context

  • Default: professional + blog
  • Reference: voice-profiles skill for definitions and auto-detection.