SnapshotJuly 2026a point-in-time write-up, left as written

Everything through Claude

A working principle that sits next to the thesis and earns its place alongside it:

Do everything through Claude. Even the things you could do faster on the side.

This is counter-intuitive at first. If you can type a command, read a file, or make a small edit in two seconds, routing it through a conversation with an AI seems like pure overhead. A few reasons to do it anyway.

Claude only learns from what it sees

If you make a change in another editor, run a command in another terminal, or make a decision in your head, Claude has no way to know. That action doesn’t feed into the context. It doesn’t teach future sessions. It doesn’t land in AGENTS.md or a Skill. It evaporates.

If you make the same change through Claude, everything is different:

Everything you do outside Claude is invisible. Everything you do through Claude compounds.

It gives you a log

Every Claude Code session is logged locally as a JSONL transcript under ~/.claude/projects/. When you want to know what you worked on last Tuesday, how much time you spent on a feature, or what the exact trajectory of a decision was, the transcripts are already there: searchable, and minable by the LLM.

Concrete uses:

Living at the edge

The arrow of history points toward working with AI all the time. The best way to learn where the edge is, and how to move it, is to use AI for everything you can, including the things it can’t quite do yet.

This matters especially for engineers. The discipline I work under: don’t write code directly unless there’s a specific reason to. When you do need to write code, go through Claude, even for the smallest things, even for things you could type faster yourself.

When Claude struggles or fails, don’t fall back to doing it manually. Ask the more useful question:

What needs to change so Claude can do this?

A better prompt. A new entry in AGENTS.md. A Skill. A different set of tools in the harness. A change to the repo’s structure so Claude can navigate it better. Every “Claude can’t quite do this” is a concrete opportunity: to shape the system, or to understand exactly where the frontier is.

People who go around Claude to avoid small friction keep their current skill set. People who push every task through Claude (even when it’s slower today) learn the system, shape the system, and end up with enormous leverage as the system improves.

A practical side effect: watching Claude work is how you learn the underlying tools and frameworks. Every tool call, every command, every library choice is visible in the transcript. You pick up the pieces by seeing them used in context. The terminal session is a running tutorial.

Where the leverage comes from

The rest of this course (building a refined concept, managing context, writing Skills, closing the feedback loop on failures) depends on this habit. If you do work outside Claude, none of those mechanisms can reach it.

Resources