Learning to crawl when everyone else is running

Everyone is running with AI. But what happens when you skip the part where you learn to crawl?

Over the past year, I have shared posts about Vibe Coding and using AI as a superpower. I have been written about building instruction files for GitHub Copilot and creating Microsoft Copilot Cowork skills. All of those posts share a common theme: AI is most powerful when you teach it how you work.

But here is the thing nobody tells you - you cannot teach AI how you work until you understand how it defaults. And understanding the defaults means slowing down enough to notice them.

This post is about what I learned when I stopped running and started paying attention. It started with a single sentence pattern I could not unsee. It ended with a set of guardrails I now use every day. The path between those two points taught me more about my own writing than years of writing without AI ever did.

The sentence that made me stop

I was reviewing a draft - one I had written with AI assistance - and a sentence caught my eye. Not because it was wrong. Because it was too familiar.

Windows 365 provides flexibility, improves security, and simplifies management.

Three items. Parallel grammar. Closed with “and.” I scrolled up. There it was again. And again. Nearly every paragraph had a sentence built the same way. Three things, neatly packaged, separated by commas.

I had not written those sentences consciously. I had accepted them. The AI proposed them, and they sounded polished enough that I moved on. But once I saw the pattern, I could not stop seeing it.

That was the moment I realized I had been running on Autopilot without knowing it.

Learning the mechanics behind the defaults

The three-comma pattern is not a quirk of one model. GPT, Claude, Gemini, Microsoft Copilot - they all do it. Once I started looking, I found the same handful of patterns appearing everywhere. Not because these models copy each other, but because they share the same optimization targets: clarity, helpfulness, and broad appeal.

The result is what I have started calling Corporate Neutral English - text that is correct, polished, and completely devoid of personality.

Beyond the three-comma sentences, I noticed AI defaults to the same transition words everywhere (“Furthermore,” “Additionally,” “Moreover”) because they are safe connectors that move text forward without committing to an opinion. It reaches for the same vocabulary (“leverage,” “robust,” “comprehensive,” “streamline”) because those words score well on professionalism without saying anything specific. And it produces paragraphs of uniform length with predictable pacing because that is what “helpful” looks like to a model trained on broad appeal.

None of these patterns are wrong individually. Humans use them too. The difference is frequency. AI does these things constantly and consistently - across every paragraph, in every response. When I went back through my recent drafts, I could see exactly where I had stopped thinking and started accepting.

Why understanding the mechanics matters

Understanding why AI writes this way is what unlocked everything that came after. The patterns are not random. They are predictable outcomes of how language models optimize output. And predictable means counterable.

Here is the key insight: you cannot write an effective instruction like “sound like me” because the model has no idea what that means. It interprets “professional tone” as Corporate Neutral English - the very thing you are trying to avoid. But you can write an instruction like “limit commas to two per sentence” or “never open with a transition word.” Those are observable behaviors. The model can follow them consistently.

That realization is what connected the dots between learning AI patterns and building effective instruction files and Microsoft Copilot Cowork skills. The instruction files I had written before were vague. They said things like “write in a conversational tone.” Now I could be precise because I knew exactly what I was countering.

From patterns to guardrails

Once I understood the mechanics, building guardrails became straightforward. Each pattern I identified translated directly into a concrete rule.

The table above looks obvious in hindsight. But I could not have written it before I understood the defaults well enough to name them. That is the crawling part - the slow, deliberate work of learning how the tool behaves before trying to control it.

Your style is not your vocabulary - it is your rhythm

When I analyzed my own writing across blog posts and customer communications, the distinctive elements were not specific words. They were structural:

• Short paragraphs, often a single sentence for emphasis
• Rhetorical questions to introduce new sections
• Real examples before abstract principles
• Opinions stated directly, then qualified afterward
• Trade-offs acknowledged rather than hidden

Most people think their style lives in vocabulary. It usually lives in rhythm. That finding came from recent style-transfer research - preserving an author’s formatting, pacing, and paragraph structure is far more effective than mimicking specific words.

If you want to build guardrails that work, start by analyzing 10-20 samples of your own writing. Look for structural patterns rather than vocabulary choices. Then map those patterns against the AI defaults. The gap between the two is where your guardrails live.

What my guardrails look like today

I turned everything I learned into concrete rules. Here is what that looks like in practice:

name: Preserve my voice

instructions:
  - Write as an experienced practitioner, not a marketing writer
  - Prefer concrete observations over abstractions
  - Reference real examples when available
  - Avoid generic introductions and conclusions
  - Avoid excessive three-item lists
  - Do not use "leverage", "delve", "transformative", "robust",
    or "comprehensive" unless genuinely necessary
  - Vary sentence length deliberately
  - Preserve opinions, uncertainty, and trade-offs
  - Sound like a knowledgeable colleague talking to another professional
  - Limit commas to two per sentence
  - Remove filler qualifiers like "It is worth noting that"

Example: Writing style instruction for AI assistants

Every rule in that list traces back to a specific pattern I identified. Limit commas to two per sentence - because of the three-comma pattern. Remove filler qualifiers - because AI defaults to empty lead-in phrases. Avoid “leverage” and “robust” - because those are statistical attractors that signal machine output. None of this is abstract preference. It is direct countermeasure.

Final thoughts

Learning to crawl means doing the slow work that everyone else is skipping. It means reading your own AI-assisted output critically enough to notice the patterns. It means understanding why those patterns exist before trying to fix them. And it means turning that understanding into guardrails specific enough that the AI can actually follow them.

AI can generate ten pages in seconds. Experience is knowing which two paragraphs actually matter - and which sentences you accepted without thinking.

This is why I decided to add an AI transparency page to this site. Not as a disclaimer, but as a statement about how I work. AI plays a role in my writing process. The ideas, the judgment, and the final responsibility remain mine. The guardrails are what keep it that way.

–Jesper

Header image attribution: Image created with help from Microsoft Copilot