I’ve been using Claude Code as my primary coding agent for a while now, and at some point it stopped being a terminal tool and started feeling like an operating system for how I work.
Here’s how I have it set up.
My CLAUDE.md has one guiding rule—simplicity first, speed always. I configured Claude to think before coding, make surgical changes, and never over-engineer. Every changed line should trace directly to what I asked for. No speculative features. No premature abstractions. I also asked it to be funny. Like TARS from Interstellar.
I run 6 plugins—Linear for issue tracking, Slack for searching channels and drafting announcements, Telegram for messaging Claude from my phone, Swift LSP for my iOS projects, Skill Creator for building custom skills, and Claude MD Management to keep project instructions clean.
The Telegram plugin is probably the wildest part of this setup. I paired my Telegram account with Claude Code, and now I can message it from my phone to navigate my filesystem, suggest blog ideas, or kick off tasks. It feels like having a remote engineering copilot in my pocket. I was on the couch when I asked Claude to suggest blog ideas for this very site—and this post is one of them.
One of my favorite workflows is the FeatureOS MCP integration. We built FeatureOS at Skcript, and it has its own MCP server that lets Claude access feature requests, changelogs, knowledge base articles, and team data directly. I have a /changelog command that pulls recent product changes, writes a changelog entry, and publishes it to FeatureOS—all without opening the dashboard. Once it’s published, I use the Slack MCP to post the changelog to the team channel so everyone knows what shipped. The whole loop—writing, publishing, notifying—happens in one conversation.
I have 35+ skills installed. A full blog engine that handles writing, rewriting, SEO auditing, schema generation, editorial calendars. Frontend design skills. SwiftUI code review. Apple-style copywriting. A project management system for solo developers. Each skill is a specialized prompt that turns Claude into a domain expert. Instead of switching between tools, I type /blog write or /seo-audit and the agent takes over.
I wrote a custom status line script that shows me the current directory, model name, context window usage as a progress bar, and git branch with file counts and line diff stats. Something like:
site-karthik | Opus 4.6 | ██████░░░░░░░░░ 40% | (main | 3 files +47 -12)Small thing, but it keeps me aware of how much context I’ve burned and what’s changed in the repo.
Permissions are wide open. Bash, all MCP tools, web fetch, search, file reads—all auto-approved. I want Claude to execute, not ask. The friction of approving every shell command kills the flow state. If I’m going to trust an agent with my codebase, I might as well trust it all the way.
Claude Code also has a persistent memory system. It remembers my preferences, project context, and corrections I’ve given across conversations. When I told it “never use find via Bash, always use the Glob tool,” it remembered that forever. This is underrated. It means Claude gets better at working with me specifically, not just at coding in general.
My typical day—morning, I check Linear issues and scan Slack. Then I pair program on whatever I’m shipping. On the go, I message Claude via Telegram when something comes to mind. For writing, I use the blog skills to draft and optimize posts. For DevOps, I manage Cloudflare, Coolify deployments, DNS, and server configs. All from the terminal, or from my phone.
The gap between thinking and doing has never been smaller for me. I don’t open a browser to check Linear. I don’t open Slack to search for a thread. I don’t open a CMS to write a blog post. Everything happens where I already am.
The future of development tooling isn’t a better IDE. It’s an agent that knows your entire stack and responds to a text message.