Claude Code Q&A: What Works, What Doesn’t, and What Will Save You Hours

 Last month, we held our inaugural Claude Code Camp, where builders from Every Studio hopped on a Zoom with Every subscribers to share how they're using Claude Code to build real products at Every—Spiral, Cora, Sparkle, and Monologue.

What unfolded was part show-and-tell, part group debugging session, and part philosophical discussion about what it means to build software alongside AI. Over two hours, we explored custom commands, parallel workflows, and the messy reality of working with agents that sometimes go rogue.

Some patterns that emerged from the conversation:

1. Planning is paramount—not as rigid doctrine, but as a way to keep Claude from wandering off into expensive tangents. Kieran Klaassen, general manager of Cora, treats plans like detailed maps for a journey, rather than instructions for how to walk.
2. Structure helps more than it constrains. Even solo developers were curious about how practices from traditional coding, like pull requests (drafts of proposed code changes), could be used in AI-assisted workflows to maintain order and ensure steady progress.
3. When it comes to working with AI, providing context from the start is essential. It’s what allows the AI to act like an informed teammate—understanding your goals and constraints—instead of a stranger dropped into the project midstream.

Our team has uncovered some cool tricks, and some principles we believe will become essential to AI-enabled coding. But we’re still experimenting, failing, and iterating in real time. And with new releases, like Claude Code’s new agents, which dropped about a week after this session, the landscape keeps shifting beneath our feet.

Want to follow along as we keep testing Claude’s capabilities? The next subscriber-only Claude Code Camp is scheduled for Aug. 22. Subscribe to Every and keep an eye on your inbox for when the invite drops.

Read the Q&A for yourself

Kieran led the session, fielding questions about his intricate GitHub-integrated Claude workflow while Dan Shipper, Every's CEO, facilitated the conversation.

The session featured a parade of hacks and workflows: custom slash commands that launch multiple AI instances to research problems in parallel, workflows that let you run three features simultaneously without conflicts, context files that teach Claude your codebase's architecture, and even an experimental logging system Kieran once tried that "fined" Claude $100 for failures. Team members demonstrated everything from using voice-to-text for faster prompting to running Claude Code on mobile phones via SSH tunnels.

Here are the highlights of the Q&A.

Planning and prompting

Before any code is written, the team focuses on planning—writing a detailed, high-level description of the work so Claude doesn’t wander into irrelevant (and costly) territory.

Q: What makes a “good plan”? You said a good plan is important but also not to have a mega-prompt. I would have expected a good plan is a mega-prompt.

Kieran Klaassen: The mega-prompt still works for a good plan. But I think the question was more about priming the model. I lean into the agentic aspect for research or coding—it’s better to give Claude a mega-prompt for the plan (so it knows every step to follow) but not for how to work. The “how” should be left flexible.

Workflow and delegation

In AI-assisted development, a pull request (PR) isn’t just for code review—it’s a checkpoint that lets the AI complete a full pass before a human steps in. This prevents wasteful mid-course corrections.

Q: How often are you correcting Claude’s work?

Kieran: It depends on the initial research and feature complexity. I’m rarely correcting mid-flight. Planning and reviewing are the most important steps.

Q: When I delegate to Claude Code, it sometimes gets a bit rogue. How do you prevent that?

Kieran: If you have a precise, detailed plan, it never goes rogue.

Technical practices

Test-driven development (TDD) flips the normal process: Write the test first, then code until the test passes. With Claude, you can build this into the plan so it codes, tests, and fixes in a loop.