N.K.

← Projects

Projects · Tooling

Working with Claude Code

I study economics, and most of my time goes to research and writing. Claude Code, a command-line coding assistant, is one of the tools I use along the way. This page is how I run it.

spec plans writes reads review reconciled
Role A tool I direct
Method Plans, specs, defined roles
In research Reference checks & review

Where it fits

Underneath my research it's a tool, no different in kind from the reference and search tools I'd reach for anyway. It has no hand in the economics or the writing. Those are mine.

How I run it

The difference is in how the tool gets run. I don't type a request and take whatever comes back. Work starts from a plan and a written spec, and the harder parts get split off to different models. It all runs against rules I set in advance, so nothing irreversible happens unchecked and the tool can't skip the reviews it owes. I also keep a retrospective check on the ways I might route around those rules myself.

Building that structure is most of the effort, and it looks more like planning research than writing code.

In the research

In the research itself I keep it on a short leash. It checks references and does the legwork around an analysis. It also runs several independent passes over a draft, each arguing against it, which I then reconcile myself. When I'm tracing something like how a central bank's signal moves a currency, that leaves me free for the argument.

research project 00_admin brief, plan, tracking 01_notes working notes 02_data sources, cleaned 03_code analysis, reproducible 07_references the literature 10_paper the manuscript (mine) reviewed before it ships
How a project is laid out

Still a tool

None of this is fixed. I change the setup constantly, and some of the rules I added I've since had to loosen when they got in the way more than they helped. It's a tool, and the research is the part that matters.