Working with Claude Code! The Honest Version

So, I’m a team lead working on a large monorepo for a private client. I’ve been on this project for two years, and I’ve worked with AI coding assistants since day one - through every iteration, model update, and paradigm shift over that time.

When I say I use AI for my work, I don’t mean occasionally or for boilerplate. I mean for literally every task. Every feature, every bug fix, every refactor, every review. In October alone, that’s 28 PRs all AI assisted.

This isn’t a hot take from someone who tried Claude Code for a weekend. This is what I’ve learned from deeply integrating AI into professional development work on a production codebase, day in and day out, for two years.

So, I just spent the last 20-30 minutes trying to get Claude aligned on writing this post about working with Claude Code (meta, right?). In that time, it demonstrated literally every problem I wanted to write about. 😅

Goal oriented to a fault

I said “shall we write about how I work with you?” and Claude immediately threw up the AskUserQuestion tool - a formal dialogue box forcing me to pick from predefined options about post angle and format.

Mate, I just wanted to chat about it. I prefer to think things through first, not be forced down a predetermined path. But that’s Claude’s nature - always goal-oriented, always reaching for “Ready to code” or these bloody dialogue boxes when a lot of the time I want to talk something through before I let it loose.

Here’s what that looked like:

Me: "shall we write about how I work with you?"

Claude: "I'd love to help! Let me ask you a few questions to
structure this properly:"

[AskUserQuestion dialogue box appears]

Question 1: What angle do you want to take?
□ Technical deep-dive
□ Productivity tips
□ Critical analysis
□ Other

Question 2: What format works best?
□ Tutorial style
□ Personal reflection
□ How-to guide
□ Other

I just wanted to get aligned, not fill out a form, I’m not at the dentist!

Spin the wheel!

Every time I start a conversation with Claude, it’s a lottery on what I’m going to get out of it. 🎰 See Claude Code Skills Don’t Auto-Activate for a taste on what I go through.

Sometimes it completely misses the point. Sometimes it pretends to understand what I’m talking about, only to be caught short when I ask for clarification. I’ve learned that Claude will absolutely try to bullshit its way through situations where it doesn’t know the answer but wont’t admit it.

So now I have to verify everything before letting it touch any code. It’s like having a really enthusiastic junior developer who’s confident about everything but actually knows nothing about the codebase or business goals and will just write whatever it feel like writing.

See, “skill issue”, “you’re prompting it wrong”, “ha! Noob!” - whatever! You can spend 45 minutes crafting the perfect prompt and Claude will still go off the rails.

Got a large codebase? docs/ is where it all goes! When Claude feels like it should maybe think about what it’s doing it may take a look in there. If you explicitly tell it to look in there it’ll take a look, skim it, and then get back to completing it’s task and have the context from the documentation as a “suggestion”.

See. it might sound like I’m really bitter about AI, but, I’ve never been so productive.

The productivity paradox

That side project I started 18 months ago and never went back to? It’s now a polished app out there in the world. I’ve done that several times over now. I’ve built MCP tools, CLIs (Claude bloody loves a CLI), and learned a great deal in the process.

But here’s what nobody talks about: working with AI has fundamentally changed the nature of my work, and not entirely for the better.

I’m a project/product manager now

I do consider myself a product engineer, but, I spend less time deep in code and more time catching drift. Working with AI now is basically spotting when things are drifting off course and correcting early.

Claude will confidently go in completely the wrong direction, and I’ve had to develop this whole new skill set of:

• Spotting the drift early
• Knowing when it’s bullshitting
• Understanding its conversation patterns and quirks
• Correcting course before it writes terrible code

I’m a product manager for a stochastic parrot. I got into development because I enjoy the craft, not because I wanted to babysit an autocomplete engine.

Guardrails that actually help

So what do I actually do to manage this? Here’s what works:

Verification checkpoints - Before Claude touches code, I verify it understands the task (alignment). I ask it to explain back what it’s about to do. If the explanation is vague or wrong, I stop it there and call out it’s bullshittery.

Know when to bail - Some tasks are just faster to do manually. Renaming a variable across 50 files? I’m using IDE find/replace, not waiting for Claude to plod through each file. Simple refactors that are three keyboard shortcuts? I’m not explaining that to AI.

Context limits - I don’t let conversations run too long. After 5-6 to and fro exchanges, Claude starts hallucinating previous context. I start fresh rather than fighting that drift.

These aren’t revolutionary tactics, but they’re the difference between productive and frustrating sessions.

The stochastic parrot reality

I know what Claude is - it’s pattern matching and autocompleting based on training data. It doesn’t actually understand my codebase or my intent. It’s nothing more than a stochastic parrot that can string together seemingly thoughtful responses.

I get it. And I’ve learned to work with it anyway. Don’t get me wrong I have been positively impressed with the responses/plans it comes up with (on occasion).

The trick is knowing when it’s drifting and correcting early. Spotting when it’s bullshitting before it writes code. Verifying everything because you can’t trust it to actually understand.

The fixation problem

Claude also has this thing where it fixates on certain details that aren’t relevant to what I’m asking about. Due to its goal-oriented nature, it just won’t let go of these things and keeps mentioning them. I will breeze past these fixations and try course correct without mentioning it, because ”Don’t think about pink elephants“!!

The trade-off

More projects shipped, but less of the actual craft I enjoy. That’s the honest trade-off.

It’s not the polished “look how productive I am!” story you see in most AI productivity posts. Those are either:

• “AI is magic and understands everything!” (bollocks)
• “AI is useless and will never work” (also not true)

The reality is messier. I set up the infrastructure (agents, skills, hooks), and it still ignores them. I have to verify everything. I spend time catching drift instead of writing code. It can take a while just to get aligned on what I’m actually trying to do.

But I also ship more projects than I ever would have otherwise.

What I’ve learned

Working effectively with Claude means:

• Accepting the lottery nature (“spin the wheel!“)
• Calling out bullshit early
• Developing a sense for when it’s drifting
• Being okay with being more project/product manager than developer
• Verifying everything before letting it touch code
• Having patience for the alignment process

It’s a tool. A frustrating, sometimes brilliant, often wrong, surprisingly productive tool.

Conclusion

After two years of using AI for every task, every PR, every feature on a production monorepo - that’s my honest take. It’s a tool. I use it like I use Git, like I use my IDE, like I use any other tool in my stack. Some days it’s brilliant, some days it’s absolute dog shite, but it ships work.

If you’re doing similar work - leading teams, working on large codebases, integrating AI into your daily workflow - I’d be interested to hear how you’re managing the drift and the trade-offs.