I'm becoming AI-blind

Who Should Read

A developer who shares Claude-drafted design docs or proposals with their team and keeps getting follow-up questions from readers

Core Mechanics

• AI-flavored documents don't sink in
  • The author says they increasingly read colleagues' documents without being able to focus on the content.
  • They end up asking the sender about things the document already covers.
• The common thread was AI traces in the text
  • A design doc carried Claude-specific analysis and phrasing.
  • A 20-page marketing deck mixed in nonsense product architecture.
  • A requirements doc described a simple concept very verbosely.
• AI prose has recognizable patterns
  • Examples include "It's not selling X, it's selling Y".
  • Pitching every small detail as a breakthrough is another giveaway.
  • The author jokes you shouldn't sell an RBAC checkbox view like you invented fire.
• The author feels 'pre-trained' on AI filler
  • After seeing many AI-generated LinkedIn posts, emails and sites, their brain spots low-effort AI content quickly.
  • It then ignores it and moves on without thinking.

Community Reactions

• Reading AI text forces the brain to supply meaning
  • One commenter says their brain does a just-in-time rewrite to give the words value, which is exhausting.
  • They say it gets worse with complex topics.
• Polished Claude learning material didn't land
  • A commenter built Romanian verb charts for their son with Claude and kept blaming their own focus.
  • Having Claude rewrite sentence by sentence for clarity made the concepts easy to understand.
• Same problem with code comments and plans
  • A commenter says Claude's PR comments are hard to parse and asks colleagues to replace them with a one-liner.
  • Others say Claude's plans and even AI code feel hard to read.
• Counterpoint: AI text can beat what humans wrote
  • A commenter says it has no typos and clear structure, and beats docs that are just lists of commands and URLs.
  • Another notes verbose requirement docs existed long before AI.

How to Apply

• Compress one AI doc by hand before sending
  • Pick a doc you will share today and put the key decisions in 3 lines at the top.
  • Delete or rewrite any part you can't summarize.
• Strip AI-sounding phrasing from docs
  • Find and remove "It's not X, it's Y" contrasts and hype wording.
  • Describe minor features plainly by what they do.
• Rewrite PR comments as one human line
  • Replace multi-line Claude comments with a one-line explanation of why the code exists.
  • Add an AI-comment cleanup item to your review checklist.
• Attach the original prompt to shared docs
  • If AI drafted the doc, paste the prompt at the end.
  • Let reviewers read the prompt first to grasp your intent.

Try It

# Paste into Claude or ChatGPT (simplified version of the article's idea: auditing AI-flavored docs)

Read the document below and find parts that are 'AI flavor without content'.

Criteria:
1. Contrast-style pitch lines like "It's not X, it's Y"
2. Small details inflated into breakthroughs
3. Technical jargon or architecture unrelated to the document's topic
4. Sentences that read like unsure internal reasoning

Document:
{{document}}

Output format:
- Problem sentence / reason / plain rewrite (as a table)
- End with a 3-line summary of the key decisions

Next thing to try: assume {{reader}} is seeing this for the first time and ask for the 3 questions they would likely ask.

Terminology

Related Papers

• AI is a tool and using it well takes skill

  A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome

• Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

  Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.

• GPT‑6 and Intelligent UI for everyone

  GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.

• Base Models Can Reason By Taking a Cue From Training Data

  Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.

• Labels Override Definitions in Jev-Style Typed Decision Models

  Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.

• How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs

  Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Related Resources

• https://cymerys.com/w/im-becoming-ai-blind
• https://spawn-queue.acm.org/doi/10.1145/3807966