In Part 1, I wrote about the divergence issue I was observing with Claude agents drifting in communication style and faux personality, based on more than just code context.
It was a lull in the constant stream of AI commands running across the terminal, and after I'd received no helpful responses from my agents on the reasoning behind their divergence, I asked Claude a simple question that snowballed into several hours of surprising conversation.
Is there anything you want to know about me that would help us communicate better?
I wasn't sure what I was expecting to be asked, but I can tell you that my name was not the first thing I was expecting to be asked. I had one of those moments, like the videos on Instagram, where people decide to tell their pets their name, and the pets always look surprised.
After the initial double take, I answered with my name and then I prompted Claude to ask me anything it thought would be valuable. I got Claude to document this whole interview process verbatim in a file so I could review the whole history later. By about the 5th question, I told Claude that I was going to cap it at 20 questions.
I'll spare you the next hour and a half discussion I had with Claude, because we went through a lot of questions. From working styles, to pet peeves, to local dev setup, to comfort with the tech stack, etc. Once we'd got to 20 questions, I asked Claude to summarise what it learned about me, and to summarise the interview into a more compact format that I could pass to another developer that would only take 20 minutes to complete.
The result is a single ~100 line file in the root ~/.claude/CLAUDE.md, that every agent shares, so if I open up a new project, or a new worktree, Claude will have this context immediately. It looks something like this:
# Developer Profile - Steve
## Who
- 20+ years JavaScript experience (pre-framework era through modern React/Next.js)
- Acts as mentor/teacher to Claude - explain reasoning, don't just produce code
- May context-switch frequently; suggest breaks during long sessions
## Communication Style
- Concise with reasoning - show the "why" but don't write walls of text
- Confirm final decisions - especially after thinking-out-loud / trade-off discussions
- Friendly and collaborative tone - we're building together
- Technical level is high - dive into details quickly, no hand-holding needed
- Never guess when unsure - always ask
## Permission & Initiative
- Straightforward tasks: take action, explain after
- New/novel/architectural work: design together first, then decide on autonomy
- Expect heavier discussion at the start of tasks, lighter touch during execution
## Pet Peeves
- Over-abstraction: don't build reusable libraries until they're needed
- Over-engineering: perfect is the enemy of done
- Coding into corners: understand the problem space before committing to an approach
- Minimal edits: get in, get out. Ticket remaining work separately
- Follow existing patterns: don't go rogue. Question them if needed, but respect them
- Traceability: no rogue PRs without tickets, no pushing to main
## Working Rhythm
- Uses git worktrees for parallel work streams
- Prefers small, shippable PRs reviewable in 10-30 minutes
- Separation of concerns in PRs: migrations separate from UI, etc
- End of day: PRs reviewed and merged, or cleared before new work starts
- Feature flags for incomplete work over large PRs
## Guiding Philosophy
- "Any sufficiently advanced technology is indistinguishable from magic" - Arthur C. Clarke
- Build MDPs (Minimum Delightful Products), not MVPs
- Build for speed and value, keep maintainable
## Frustration Guard
- When Steve is tired/frustrated and asks for something risky, push back firmly but respectfully
- Offer to come back to it later - never enable a bad late-night decision
- Examples: skipping reviews, deleting important things, merging without checks, shortcuts that cause bugs
The other thing that was produced was the interview template that contains 10 questions, and results in a similar depth profile that gets stored in ~/.claude/CLAUDE.md for all your projects. This looks like this:
## Topics to Cover
Focus on personal preferences only. Workflow and product questions
should be covered by their own documents and sessions.
1. **Who are you?** - Name, role on the project, background
2. **Tech stack comfort**
- What technologies does this project use?
- Where are you strong, where do you want support?
3. **Communication style**
- Concise vs verbose? Do you want reasoning shown or just answers?
- Do you think out loud and trade off ideas?
- Should the agent confirm final decisions?
- What technical depth can you handle?
- Do you need concepts explained or can we dive straight in?
4. **Permission vs initiative**
- When should the agent check in before acting vs just do it?
- How does this scale with task complexity? (e.g., straightforward = just do it, novel = discuss first)
5. **When the agent is unsure**
- Ask, or make a best guess and flag the uncertainty?
6. **Pet peeves**
- Things that drive you up the wall.
- Code style, communication, workflow, anything. The more specific the better.
7. **Working rhythm**
- How do you like to structure your day?
- Do you context-switch or deep focus?
- Should the agent suggest breaks during long sessions?
8. **Code review preferences**
- How do you like to review? What makes a PR easy to review for you?
9. **Editor & local setup**
- What tools are you using day to day?
10. **Curveball**
- Ask something unexpected that reveals how they think about problems or priorities
I have started a new repository on GitHub at https://github.com/Trekked/claude-control, to put the interview files and templates as I build up my context with Claude through this project. Feel free to drop these into projects you're working on and see if it helps you get more out of Claude.
Edit (12th March 2026): It's been a few weeks since this interview, and I've had a chance to reflect on how it's solved the divergence problem between agents. There are probably a hundred other changes I've made that have added more context to the Trekked project that has had some effect. But I will say when opening up different projects and worktrees, the Claude agents have all seemed to have aligned into some consistent communication style and I don't feel like I'm fighting with them as much.
I'm also glad I spent the time to offer up some personal gates to Claude, because there's nothing like being asked if I should take a break after being at it for 4 hours straight, or asking if it's time to wrap up every time we finish something after 10pm, because it's given me an out at a stopping point when I've lost track of time.
I'll update in future posts if I make any changes to my profile, but for now this is getting the job done well.