Working with AI Coding Tools

A guide series for librarians and researchers working with AI coding tools.

These guides cover the kind of work librarians and researchers do with AI coding tools: cleaning a dataset, screening a literature set, building a small visualization or static site, setting up some recurring reporting.

The chapters are written so you can dip in wherever's useful. The examples use Claude Code, but the methods apply to other tools too.

The Guides

Your First Hour with Claude Code

Setting up Claude Code and running a first session on a folder of your own, with attention to the interface and the approval flow.

Git in Practice

When an agent edits many files in a single session, you'll want to be able to review the changes and undo any of them. Git is the standard tool for that.

How to Talk to the Tool

How to write a workable prompt and refine it over several turns of back-and-forth with the tool.

What People Build

Examples of work that AI coding tools handle well, and how to tell whether yours will fit.

Thinking in Steps

Splitting a larger task into stages so each one can be inspected on its own, before moving to the next.

When the Vibes Go Wrong

Common ways AI-generated code fails, often while the agent is confidently reporting success.

Building Skills for Claude Code

Skills let you save the instructions and edge cases that keep coming up across sessions.

Verifiable Goals

When a goal can be checked automatically, the agent can iterate on its own. When it can't, you have to stay in the loop.

Working with Tabular Data

Cleaning spreadsheets and CSVs while checking the tool's assumptions about columns, types, and structure.

Working with APIs and External Data

Working with an API and what to do when the responses don't match what you expected.

Your First Web App

Most small web projects don't need a server. When yours does, how to add one without expanding the project unnecessarily.

Row-by-Row Dataset Processing

Methods for datasets where every row needs a separate judgment call (screening literature, coding open-ended responses), with focus on making the work reviewable later.

Web Scraping as a Last Resort

Before scraping a page, look for the API it's quietly using in the background. The API is usually easier to work with.

Project Context with CLAUDE.md

CLAUDE.md holds your durable project context. Keep it selective and prune it before it gets too long to read.

Working with Existing Codebases

When you inherit a codebase you didn't write, the first job is reading it carefully before changing anything.

What retrieval-augmented generation is and when it helps. For some questions a direct tool-based lookup does the same job.

Orchestrating the Browser

Comparing Playwright, MCP browser tools, and computer-use loops for work that has to drive a browser.

Bonus

Example project

The Rise of Open Access

A small worked example, starting from a research question and ending in a chart, with each step kept short enough to inspect.