→ what they shipped
→ what they wrote
→ what they could explain

a fourth axis has appeared, and the people who don't see it yet are working a layer below the people who do.

It's is what your AI can do, configured by you.

the .claude directory is the receipt.

Before 2026, AI usage was a function of the model. you had access to claude or you didn't. you had access to GPT or you didn't. the variance between two developers using the same model was small — both pasted prompts, both got outputs, both moved on.

that's gone now.

two developers running the same model can produce wildly different work, because the model isn't the input anymore. the configuration is.

the configuration lives in a folder. usually .claude/ in the project root. sometimes spread across a workspace. either way, it's now visible, portable, and increasingly - readable.

the .claude folder is small. but every file in it does work.

→ claude.md
the project context. tells the model what this codebase is, what conventions it follows, what's idiomatic for this team. the model reads it before responding.

→ /skills/
modular instruction sets. each subfolder has a SKILL.md that defines a discipline (data analysis, frontend design, document writing). the model loads them on demand based on what the task needs.

→ /agents/
subagent definitions. instruction sets for specialized model instances that handle defined slices of work (a tester, a reviewer, a planner).

→ MCP wiring
not inside the folder, but tied to it. MCPs connect the model to live systems: the filesystem, a vault, a task tracker, an analytics tool. the model can read and write across them.

most developers know one of these exists. fewer use two. very few have configured all four.

a resume tells you what someone did. a .claude folder tells you how they think.

you can read someone's .claude folder in 90 seconds and know more about how they operate than from an hour-long interview.

it's also the receipt of effort. configuring an AI well takes hours. the person who's done it for their workflow has demonstrated they can do it for yours.

if you have nothing yet, build in this sequence. each one is more leverage than the last, but harder to do well.

→ project context — easy. a claude.md file with the codebase's tech stack, conventions, and the kind of help you want. start here.

→ skills — medium. modular .md files per skill. each one tells the model when to apply specific knowledge. you only need 2-3 to feel the difference.

→ subagents — hard. defining specialized instances that operate semi-autonomously requires you to know what divides cleanly and what doesn't. start with one (a tester or reviewer) and grow from there.

→ MCP wiring — system-level. connect the model to your filesystem, your task tracker, your notes. this changes what "asking your AI" even means — the model isn't a chatbot anymore, it's an operator inside your existing tools.

the verb shift is the real story.

before: you ask the model a question. you get a response. you continue or you don't.

now: you give the model a job. you specify what done looks like. it runs until satisfied or escalates.

asking and scoping look similar from the outside. they're not. asking is a single-turn negotiation. scoping is a contract.

high-output developers in 2026 have stopped asking. they're scoping.

the .claude directory is what makes scoping possible. without it, every job starts from cold context. with it, the model starts inside the work.

if you don't have a .claude folder yet:

→ create one. a single claude.md file with three sections: stack, conventions, what you want from the model. fifteen minutes.

→ add one skill. pick the discipline you reach for most often (writing, debugging, refactoring). define when to use it, what outputs you want, what to avoid. another fifteen minutes.

→ pick one MCP to wire. the highest-leverage default is filesystem — the model can now read your actual files, not just what you paste. if you use obsidian, wire the obsidian MCP. if you use linear, wire that.

three pieces. forty-five minutes. the difference between your AI tomorrow and your AI today.

if you have one already:

→ read it again. when did you last update claude.md? does it reflect the project as it is now, or as it was three months ago?

→ audit your skills. which ones did you load and forget? prune.

→ add a subagent. the lowest-friction one to start: a reviewer instance whose only job is to read what you just wrote and surface what's wrong.

Read someone's .claude folder. You'll know how they think.

build yours. eventually someone will read it.

more at my Github repo:

⚙️

🧡

Here’s a quick gift for you.
Grab this guide for free to make your profile better

Use code FREE to get it for FREE!

Everything here is documented as it runs.

If this is useful, hit ⭐ — it helps others find it.

Book a 1-hour Content & Branding Strategy Call with me.

We'll look at what you're currently doing, figure out where the friction is, and map out a system that works with how you actually think and create.

Check out some of the other services I offer:

Don’t forget to reply and let me know what you thought of this newsletter.

See you on the next one
- AUNY 🧡