Claude Code is attracting a lot of press buzz and industry adoption. In this article I use Claude Code to develop new functionality and get a sense of it’s capability.

You may have seen a number of recent headlines in the financial press on Claude Code, for example the WSJ has “Claude is Taking the AI World by Storm and Even Non-Nerds are Blown Away“, which has the tagline:

Developers and hobbyists are comparing the viral moment for Anthropic’s Claude Code to the launch of generative AI

Is the buzz justified? To try and answer that, I spent a few days using Claude Code and what follows is my overview of the experience.

Getting Started

From my VSCode IDE, I can open a folder or clone a Git Repository and then right next to the Explorer editor, I have a Claude Code panel, making it super simple to start using it.

The obvious first prompt is:

I need an overview on what this code does

Claude expanded this prompt to more specific instructions as below:

"Explore this codebase thoroughly and provide a comprehensive overview of what this project does. Look at: The project structure and main directories. Package.json or similar config files to understand dependencies and scripts. Main entry points and key source files. Any README or documentation files. The overall architecture and purpose of the application. Provide a detailed summary of: - What the application does (its purpose) - The tech stack used - Key components/modules and how they relate - How the project is organized."

And returned with a very reasonable and comprehensive overview of the code. I won’t bore you with the verbose specifics of that in this article.

My Task

What I really want is to add new functionality to my application and for this to be a sufficiently complex task, so I get a good sense of the process and the outcome.

Rather than directly specifying what I wanted and letting Claude Code start writing code, I asked for a plan, much as you would do when working with a software engineer.

I want to create a new tab view under bank holding company helpers, named Adhoc Query, the purpose is for a user to type in natural language questions on the BHC data.For this we need a selector for the LLM model to be used, a text entry box for the user to type in their question and below this a text box for the response. As the user will not know the database schema, use code in the comparative_analysis_view, institution_deep_dive_view and measure_search_view to work out the schema and SQL needed to reply to the user's questions. In addition the user may not enter the full name for a institution, just part of the name, so need to cope with that in the query. Also the user will not know specific item names, so use schedules, sections and item descriptions to figure out what data they are after. Do not write any code for now, first work out a plan for me.

My requirement for the task, with brief guidance and ending with “do not write any code, first work out a plan”.

As you will appreciate on reading the above, the task is far from simple both in terms of context on the existing code base and the new functionality required.

Iterating on the Plan

Claude Code started reading code, I could see progress in the panel and at the end of this, it came up the a plan that explained the approach and files it wanted to create or edit.

I could then ask questions to get more elaboration on specific aspects and request changes or provide more guidance.

Resulting in a revised plan and more back and forth iteration, until I was satisfied with the outcome and it was time to allow code changes.

Writing Code

Claude Code has read and write access to the code in my local folder and as it starts it’s work it clearly shows what it is changing.

On completion, I can scan the full list of changes and spend time reviewing as much or as little of the new code as I wish.

It is then easy to run the application, test the functionality and ask for a change.

This back and forth, seems entirely natural and how I would code myself, think of a change, edit the code, test the change, not quite as I wanted, edit code again, re-test, repeat to completion.

The only difference is that Claude writes the code, saving me time and thinking effort to recall specific knowledge of programming language syntax, design patterns and methods/packages to use.

Let’s look at a short snippet from one of these iterations.

I entered an issue found in my testing as below:

The Data table, when I sort columns for item_value or change, the sort seems to do an alphanumeric sort, these values are numeric, so need a numeric sort

Claude Code responded with:

Clear diagnosis of the problem and then proceeded to read and edit the python file that needed changing, allowing me to view the lines that were changed.

On completion, I tested it all worked as expected.

Very good.

To Do Lists

A nice feature as Claude shows what it is doing is To Do Lists of tasks which are shown and updated for progress (as below).

Long Context

As you may imagine, the iteration back and forth to refine features or fix issues found in testing, leads to very long context windows.

To handle this, Claude at times pauses and decides to summarise the previous conversation:

This summary has a lot of information to cover the important parts of the preceding conversation, including code snippets.

Not one for me to read, but it works well as it allows Claude to continue working on my subsequent request, without losing context.

UI Layout Changes

A few times, I wanted an improvement to the UI and asked Claude for suggestions, which it proposed as Option A, B and C all within text. On my replying that I was not sure what Option B would look like, I got back a line diagram.

Simple, but effective in helping me understand if I preferred this option for the new layout.

System Stops

The most annoying feature was Claude System stops when I hit message limits imposed by by Claude Pro plan. It seemed to only take a couple of hours of work, before I was stopped and asked to wait 3 or 5 hours ( I forget which) before continuing. Needless to say that was it for the day and more than frustrating when deep into the process.

Consequently I could only spend 2 to 3 hours each day, before being stopped, which meant it took me 4 days to get code that I was willing to check into GitHub.

An overall 8-12 hours of work is not much effort for a satisfying outcome.

Daily Summary

A nice feature, is the below response from Claude to my usual “that’s all for today” – a summary of the improvements that day.

My Thoughts

I am happy with the end result, a new Adhoc Query View for Bank Holding Company data, which allows a user to ask questions in plain english.

I would not have been able to build this without Claude Code.

Working with Claude Code was easy, felt natural (except for the system stops!) and the python code is well written and well documented.

I was able to:

• iteratively improve the functionality as I tested the application and worked out what else I wanted,
• inspect the code and ask Claude to improve or simplify assumptions,
• test, get bug fixes and re-test quickly,
• utilise Claude to narrow what to re-test, given code paths,
• not worry about specific syntax in using a specific language or package feature, and
• remain focused on the functional outcome

All in all a successful outcome and process.

I will definitely be using Claude Code for the next enhancement or bug fix.

In Summary

Claude Code is empowering for product managers, analysts and developers.

It feels natural to use and offers gains in productivity.

Allowing us to prototype, add features and fix bugs.

Staying in control but relying on Claude to write the code.

Subscribe below to follow our journey as we use GenAI tools.