Why your Claude setup sucks and how to fix it

There’s really just one difference between Claude pro users and Claude noob users, and it comes down to the setup.

Most people building Claude skills spend their time perfecting the prompt, eg tweaking the structure, iterating on the output format, testing edge cases, etc. The skill gets better, yet the output stays mediocre. I’ve watched this happen across teams and it’s one of those problems that looks like a craft problem until you realise it’s an architecture problem.

The reason is simple: every prompt forces the skill to start from zero. That means it’s got no memory of your audience, no knowledge of your positioning, no understanding of how you or your team actually work. Claude is smart, but it can only work with what you give it in the moment, and “in the moment” is almost never enough.

This is a silo problem. Every skill knows its own world, with no cross-interaction and no shared foundation. It produces really recognisable failures: copy that writes about your audience instead of for them, positioning that sounds like every other B2B company in your space, content that’s technically on-brand but has none of the texture that makes your real stuff land, and half-baked PRDs that don’t quite hit the mark.

Three things that specifically break without a shared context layer:

Context amnesia. Every new session, you’re re-explaining the same things. Your ICP, your voice, your market position. The quality of output ends up depending entirely on how much the user remembers to include on any given day, which is wildly inconsistent.

Config drift. Hardcoded values go stale. Names change, frameworks update, team structures shift. At one skill, it’s manageable. At twenty skills across three teams, it becomes a maintenance job nobody signed up for. (I’m calling it now: AI Ops is coming as a job title, and this is why.)

No learning loop. When a session goes well, or when Claude makes a mistake that reveals a real gap in a skill, that insight evaporates at the end of the session. There’s no mechanism to capture it, no way for it to improve the next run, no compounding. Every session is equally good, which means every session is equally mediocre.

The fix is a three-layer architecture. A shared knowledge base that lives in your Claude Project (positioning, ICP, brand guidelines, etc) that every skill checks before asking the user for anything. A personal MEMORY.md file per user that Claude writes to over time, building a picture of how that person works, their preferences, their decisions. Add an improvement loop that silently logs patterns and gaps across sessions so the humans managing the system have real material to work with.

The whole setup takes an afternoon. Each user onboards in about two minutes. And the compounding effect over six months is genuinely significant (richer shared context, smarter personal memory, better-calibrated skills) compared to a team that spent the same time maintaining isolated skills with hardcoded values.

I wrote up the full architecture, including exactly how to set it up step by step, over here: Your AI skills are silos. Here’s how to fix that.

If you’re building with Claude at your company and want to think through how this applies to your specific setup, just reply. The patterns are transferable to any team, any workflow, and any industry.

Andrea