SE: Six Months with Claude Code
← All posts
PC<br>PyCharm
python.py<br> agent_lib<br> commands<br> plugins<br> skills
python.py<br>utils.py
12<br>13<br>14
def test_function():<br>def __init__(self):<br>print("Received hello")<br>def _close(self):<br>out = "Writing code is neat"<br>return _.close(cli)
agent-cli-plugin
Claude Code<br>Ready to implement your Jira tickets.
CLAUDE<br>Reviewing the implementation plan now.
Type a message...
Jun 24, 2026 · 5 min read
Six Months with Claude Code: How My Developer Workflow Changed
Since October 2025, I have been using Claude Code extensively for work. In the beginning, there was definitely<br>a learning curve, and I had to adjust my workflow around it. I remember back when I worked as a Kotlin developer<br>at my previous company, I used GitHub Copilot primarily for autocomplete in IntelliJ IDEA. Later on, I used<br>ChatGPT to generate small code blocks and scripts, but most of the time, I was still coding manually by myself.
But with the release of Claude Opus, everything changed. The model was so impressive that, little by little, I<br>started using it for brainstorming, researching, breaking down initiatives into smaller tasks, creating and<br>refining Jira tickets, implementing and testing code, automatically opening PRs, and even handling code reviews.
However, one thing that really bothered me was the lack of a unified IDE where I could seamlessly use Anthropic<br>models. I tried Cursor, Zed, and Claude Code in the terminal, but as a Python developer, I found myself<br>constantly switching back and forth between PyCharm and other tools. Maybe it's because I'm so used to JetBrains<br>IDEs and compared to PyCharm, other editors felt quite inadequate.
Having previously developed plugins and even a whole custom distribution of IntelliJ IDEA at a former company, I<br>already knew how to write an IntelliJ plugin. So, I embarked on writing the<br>agent-cli-plugin.<br>With Claude Code acting as my "code monkey," it was incredibly fast to come up with an MVP in just a few hours.<br>I've added functionalities and extended it support other CLI agents over time, but now I use it as my primary<br>tool to interact with Claude Code inside PyCharm. My main goal was to get proper rendering, and in my opinion, it<br>looks way better than running Claude Code in PyCharm's default terminal.
When it comes to my day-to-day processes, I use Claude Code for almost every aspect of my work. Recently, I've<br>settled into the following workflow:
Brainstorm & Research
I gather all requirements for a given initiative and create a Markdown file with all the<br>information. This file ends up being a mix of LLM generated content and manual adjustments, serving as the<br>core knowledge base for the entire initiative.
Breakdown & Ticket Creation
I break down the initiative into smaller, implementation and testing ready chunks based on<br>the codebase, saving them as smaller Jira tickets in another Markdown file. I've built a custom Claude<br>skill to format these tickets exactly how I need them. After reviewing and refining them one by one via<br>Claude Code, I use another skill to publish them directly to Jira via the Atlassian MCP.
Team Alignment
I present the tickets to my team for further refinement and estimation before pulling them<br>into the sprint.
Implementation Planning
Once I pick up a ticket, a dedicated Claude Code skill fetches it and generates a detailed<br>implementation plan in a Markdown file. Because the Jira tickets are so well specified, the model gets the<br>plan right most of the time, and I rarely have to make any modifications.
Sandboxed Execution
I execute the plan in a sandboxed Claude Code session. I created a small<br>claude-crate<br>Docker image containing bare minimal tools and Claude Code, mounting the project directory at container<br>startup. Inside this container, Claude Code runs with the --dangerously-skip-permissions flag.<br>I use my agent-cli-plugin within PyCharm to launch this sandbox. Because the host machine's<br>codebase is mounted, I can watch file changes happen in real time inside PyCharm. The best part is that<br>while it implements code, writes tests, and finishes the task, I can focus on something else without Claude<br>constantly interrupting me for permissions. Any destructive changes are isolated to the container, and<br>worst case scenario, I can just do a clean git checkout. Note that in this sandbox mode, no MCPs or extra<br>skills are configured. It's a purely bare bones Claude Code session.
Review & PR
Once the implementation is done, I use another skill to run a code review on the changes<br>made on the host machine. If tweaks are needed, I use the same session I used for the implementation plan<br>on the host. Finally, I spin up a PR using a skill powered by the GitHub MCP.
Beyond this core workflow, I have various skills for testing, PR code reviews, debugging alerts, and more. With<br>all of this in place, I feel like my productivity has easily doubled. Recently, I even built an experimental<br>personal assistant agent named Jarvis....