Most people open Claude Code and start typing. Here's what I do instead.

I've spent the past few months using Claude Code for my AI coding projects. I started out with the basic Claude subscription. Then I quickly upgraded to Claude Pro and then Claude Max. I've spent several thousand dollars on Claude Code over the past few months, including my API bills for my apps!

Some of my AI coding projects include:

• Wheelhouse: An app for trading options as part of the wheel (paid for itself after a few weeks!)
• The Operator: An app for managing training for triathlons in Ireland (fun hobby coding project)
• Vendors.ie: An Irish business intelligence site (potential new biz)

While building these projects, I made many coding mistakes, but I also learned a few things about refining my AI coding workflow.

Run /init

When firing up a new AI project, start with /init. Claude creates a CLAUDE.md file that you (or the AI) can populate with project instructions and guidelines. Claude reads this file automatically every time you start a new session for that project. No more tangents or re-explaining your codebase.

If you don't know what to put in this file? Just get Claude to ask you a couple of questions about your goals and what you're trying to do. It'll help you write.

Pro tip: CLAUDE.md works at three levels — ~/.claude/CLAUDE.md (global, all projects), project root CLAUDE.md (per project), and .claude/settings.json for permissions.

Spend 5 Minutes In Plan Mode

It's tempting to start AI coding straight away, but you'll save yourself a few hours if you spend at least five minutes in plan mode before coding.

Before I started doing this, Claude Code kept rambling time-consuming, and pricey token-burning tangents. I spent hours unravelling a mess while building a tool on Vendors.ie.

Now I start every AI coding session by tabbing over to plan mode.

After a brief chat, Claude presents a markdown doc of exactly what it's going to do, which I review before sending the agent to work. Think of it as a PRD before every coding session.

Turn Auto Accept Edits On

In a 1995 Simpsons episode, Homer wants to qualify for a work-from-disability programme so he can stay on the couch. So, he rigs a drinking bird to repeatedly tap the Y key on his keyboard to answer "yes" to every safety prompt at the power plant.

The bird works fine until it falls over, and the plant nearly melts down... kind of like AI coding a disaster.

If you find yourself tapping "Accept" over and over and over, just keep tapping until you turn auto-accept edits on. Once Claude Code has a clear remit, it can quickly work through everything you wanted to do without you having to tap Enter or 1 like a drinking bird.

Skip Permissions (Dangerous Mode)

If you find yourself stuck in auto-accept mode for longer than you'd like, try this:

Fire up a new instance of Claude Code with claude --dangerously-skip-permissions.

Be warned: with great power comes great responsibility.

Claude automatically accepts all edits, bash commands, and file changes without asking for permission. Be warned, Claude will do exactly what you tell it, including destructive stuff.

Compact Your Context

The context window fills up fast in long sessions. When it's getting full, use /compact to compress the conversation. Claude summarises everything and continues with a refined, shorter context.

You get more runway without losing the thread. If you want to start up a new chat with Claude without actually exiting and restarting, simply use the /clear command, and Claude will forget all of its context.

Save Repeatable Workflows as Skills

If you find yourself prompting and reprompting Claude, ask it to save your chat and the output as a skill. Then, you can invoke the entire workflow with a single command.

For example, lately I've been drafting my content using the app Wispr Flow. It's much faster than typing, and it's nearly 99% accurate. However, I still want to tidy up my dictated drafts and copy them to my clipboard for posting.

When I explained my problem to Claude Code, it turned my chat into an edit skill I use almost daily.

Mine Your Coding Insights

Self-knowledge is power. The /insights command reviews all your sessions from the last 7 days and generates an HTML report. What you worked on, where you wasted time, mistakes you kept making.

Mine told me a single 6-hour session is far less effective than starting a new chat for each job. It even suggested updates to my CLAUDE.md file and spotted automation opportunities in my coding workflow. I run this once per week, then copy the output back into Claude Code and ask it to review my skills and improve my CLAUDE.md file. It's also helped me learn a lot more about prompt engineering as I build.

Check Your Spend with /cost

Use /cost to see how much your current session has burned in tokens. This is useful when you're deep in a build and want to know if that refactor was worth it.

After using this a few times, Claude gave me some suggestions on how to cut my bill. For example, it helped me cache a system prompt for one of my apps and cut API costs by 80%.

Switch Models with /model

You can easily switch models mid-session. Plan with a cheaper model, then switch to Opus for the heavy lifting. There's no need to burn expensive tokens on architecture discussions.

Fire Up Multiple Instances

When working on a project with Claude, I open four or five windows in Warp (my terminal app of choice). I use one window per job. I give each window a name and use tabs to check what my agents are doing. Each agent stays focused because its context is clean.

Harness the Swarm (Sub-Agents)

Claude Opus can spawn sub-agents that work on your projects. One for UX, one for copy, one for debugging, one for frontend QA. The sub-agents report back to the master agent, which analyses all their feedback and gets to work. It's like employing a senior dev with a team.

A good swarm will cost you a lot of tokens, and you'll hit a session usage limit quickly, but the output quality and speed to production are meaningful. I upgraded to Claude Max after seeing how powerful swarms are.

Teleport Your Session

I usually run Claude Code on my MacBook, but when I'm on the go, I hit /teleport. It sends the coding session to claude.ai in a browser. I can continue working on my phone, then teleport back to my Mac later.

Connect External Tools with MCP

Claude Code can connect to GitHub, Slack, databases, and other tools via MCP (Model Context Protocol) servers. If you're running multiple agents on a real project, this is a big unlock — the agent can read issues, check PRs, query your database, and post updates without leaving the terminal.

I set up MCPs for Kit, Google Calendar, Apple Reminders and more. That last one is fun. If I type remind me in Claude Code, it automatically adds a To Do to my Reminders list.

Build Systems with Claude Design

Claude Design is super powerful and also super heavy on token usage. It's superb for creating a design system that you can quickly and easily import into Claude Code.

Come up with a few ideas for what you want your app or website to look like, taking screenshots and noting the style in question. Ideate directly in the Claude Chat.

Once you have an idea of what your app will look like, head over to Claude Design and ask them to create a complete design system.

It'll ask you a series of questions, and it can take 10-15 minutes to create a design system. When it's done, you can import it into Claude Code and iterate.

I've also experimented with using Claude Design to create presentations for an MBA that I'm doing. No more messing around with breakpoints, line spacing, and capitalisation in PowerPoint or Google Slides.

---

Get started with AI coding using Lovable.

Need an AI writing assistant? Try Grammarly with this 25% discount.