A coworker asked me how to get started with AI tooling, and I realised my answer might be useful to others. Here's roughly how I think about it as of early 2026.
𝗧𝗲𝗿𝗺𝗶𝗻𝗮𝗹-𝗯𝗮𝘀𝗲𝗱 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 (𝗽𝗶𝗰𝗸 𝗼𝗻𝗲)
At a bare minimum, ditch VS Code Copilot and get OpenCode. You can log into Copilot through OpenCode and you get a genuinely decent AI agent harness without all the VS Code bloat. Otherwise, install Codex CLI or Claude Code, whichever you have access to. If you go with Claude Code, look into the skills and extensions available.
No matter which CLI tool you use, investigate the PRD skill from the Ralph Wingham loop (https://github.com/snarktank/ralph). It's really cool.
𝗚𝗶𝘁 𝘄𝗼𝗿𝗸𝘁𝗿𝗲𝗲𝘀
Spend some time learning git worktrees. They are a lifesaver when working with long-running agents in terminals. Write yourself a script that lets you create or switch to a worktree quickly. Here is one that I hacked together that makes it abundantly easier than the existing git worktree commandline: (https://github.com/vanonselenp/dotfiles/blob/main/gz.zsh).
𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗴𝗮𝘁𝗵𝗲𝗿𝗶𝗻𝗴 (𝘁𝗵𝗶𝘀 𝗶𝘀 𝘁𝗵𝗲 𝗯𝗶𝗴 𝗼𝗻𝗲)
If you have access to ChatGPT or Claude.ai, connect every internal tool and system your organisation offers. These platforms can scan across Google Drive, Confluence, Jira, and other sources to assemble the context you need when working on something new. It's a massive time saver.
Also, if you're in meetings with Gemini (or any tool that supports it), keep your transcripts. They're incredibly useful for generating additional context later.
𝗢𝗻𝗲 𝗹𝗮𝘀𝘁 𝘁𝗶𝗽
Skip the GitHub MCP. Your agent already knows how to use the GitHub CLI, and it won't pollute your context window the way MCP integrations tend to.
That should be enough to get started. Happy to answer questions in the comments.
PS: if you happen to find a good cli for accessing Confluence and Jira, let me know!
Top comments (0)