For the past year, the loudest advice in AI circles has been simple: feed the model more. More context files, more skills, more hooks, more rules. Build a giant brain folder and Claude will finally “get” you. I bought into it too. I had a whole stack of markdown files that I thought made my prompts smarter.
Then I came across a post from an AI professional who argues the opposite. The advice is to stop copying what influencers do and start copying what Anthropic’s own team does. And what they do is surprising.
The Myth: More Context Equals Better Output
The popular belief goes like this. The more you tell Claude about yourself, your style, and your workflow, the better it performs. So people build elaborate setups with an about-me file, an anti-AI writing guide, folder protocols, and a long list of custom skills.
The author used to recommend exactly that. Back in March, their advice was this:
“With Cowork, the entire ballgame is text files.”
The setup they described was one big folder with 4 subfolders. Inside sat an about-me .md file, an anti-AI writing file, and a folder protocol pasted into the instructions. I respect this a lot, because the creator openly admits it was good advice at the time, with past models. Their message now is blunt: “do not follow this advice.”
That kind of honesty is rare. Most people double down on what they taught last season. This LinkedIn creator did the opposite and showed the receipts for why.
The Evidence: Anthropic Says So Itself
The counter-argument doesn’t rest on opinion. It rests on two pieces of evidence straight from the source.
Evidence #1: The person who built Claude Code. Boris Cherny, the guy who built Claude Code, said this:
“Every 6 months, delete your Claude MD. Delete your skills. Delete your hooks. See what the model does, and it might surprise you.”
Evidence #2: Anthropic’s own prompting guide. According to the post’s author, the guide warns that skills written for older models “can degrade output quality.”
Read that again. The instructions you wrote to help an older model can actively hurt a newer one. Older models needed hand-holding. Newer models are smarter at reading intent, so all that scaffolding turns into noise. Worse, it can steer the model toward outdated habits you wrote rules to fix months ago.
Why Overloading Claude Backfires
The expert uses an analogy that clicked for me instantly. It’s like briefing a consultant with your whole CV before every single meeting. They stop thinking about the problem. They can’t stop thinking about you.
That’s what a bloated context setup does. Claude spends its attention on your preferences, your banned words, and your folder rules. The actual task you asked about gets a smaller slice of focus. You end up with output that ticks every box in your style guide but misses the point of the request.
Here’s a quick way to spot whether this is happening to you:
- Generic answers despite heavy setup: You have pages of instructions, but responses still feel off-target.
- Rule-following over thinking: Claude obeys your formatting rules perfectly while misreading the actual goal.
- Long, wandering sessions: The model churns for ages before landing on something usable.
So What Do You Give Claude Instead?
This is where the post gets practical. The original poster breaks the new approach into three clear moves:
- What to delete (and where it hides).
- How to write one clear goal.
- The one line to add at the end of every prompt.
Step one is a cleanup. Old context hides in more places than you think. Check your Claude MD files, custom skills, hooks, project instructions, and any “protocol” text you pasted into settings months ago. If you wrote it for an older model, it’s a candidate for the bin.
Step two is clarity. Instead of ten files describing who you are, write one clear goal for the task in front of you. What do you want, and what does “done” look like? That single sentence often does more work than a whole folder of background.
Step three is my favorite part. The author tested a simple trick: ask Claude to repeat your goal back before it does anything. If it misunderstood you, you find out right away. Not after a 30-minute “Recombobulating” session that produces the wrong thing.
If you want to try it today, something like this works as a closing line (my own wording, inspired by the post):
“Before you start, restate my goal in one sentence so I can confirm you understood it.”
It’s tiny, but it saves so much wasted time. You catch misunderstandings at the cheapest possible moment.
How to Run Your Own Cleanup This Weekend
The person who shared it suggests setting aside about 10 minutes this weekend. Your old folder is probably still sitting on your desktop, so here’s a simple way to approach it:
- Back it up first: Move your old files into an archive folder instead of hard-deleting them. You can always bring one back.
- Strip everything out: Remove your Claude MD, skills, and hooks, exactly as Boris Cherny suggests.
- Run your usual tasks: Use your normal prompts with just one clear goal and the repeat-back line.
- Compare honestly: Notice where the output got better and where it slipped. Only restore what clearly earns its place.
- Set a reminder: Repeat this every 6 months, since models keep changing.
Why This Matters Beyond Claude
I think this lesson applies to every AI tool you use. Prompting habits age fast. The tricks that worked on last year’s models often become dead weight on this year’s. The smartest move isn’t building a bigger system. It’s regularly testing whether your system still helps at all.
What I love about this post is the mindset shift. Less setup, clearer goals, and a quick sanity check before the model starts working. It’s simpler, faster, and backed by the people who actually build these tools.
Check out the full LinkedIn post for the author’s complete walkthrough and more detail on each step. Then go empty that folder.