My Claude setup: Superpowers, claude-mem, and rtk

A couple of weeks ago I wrote about how I went from refusing to use AI coding tools to actually enjoying working with them. That post was mostly about mindset: learning to plan first, to stay in the loop, to treat the AI like a fast collaborator rather than a vending machine for code.

This one is about the what. A few tools and plugins have stuck around long enough that they’ve become part of how I work every day. None of them are magic (well, sometimes claude-mem feels like it is). What they have in common is that they fit into a workflow where I’m still the one steering, rather than trying to take the wheel from me.

There are three: Superpowers, claude-mem, and rtk.

Superpowers

Superpowers is the one that changed the most about how I work.

What it does is stop me from firing off a prompt and hoping. Instead, it nudges me through a structured workflow.

1. Brainstorm the idea first.
2. Then write a plan.
3. Then, and only then, implement it.

I know that sounds like the thing I already said I’d landed on in the last post, and it is. The difference is that Superpowers makes the discipline the default instead of something I have to remember to do. When I start a piece of work, it pulls me into a brainstorming conversation before any code gets written. We poke at the idea, I get asked questions one at a time, we settle on an approach, and out comes a plan I actually read. It also forces me to actually understand the domain. More than once, the brainstorming step has sent me back to a teammate to nail down a detail of the feature before I could move forward.

The results are a lot better, and the part I didn’t expect is how it made me feel. Going through brainstorm → plan → implement, I feel like I’m part of the solution again, not a spectator watching a diff scroll past. I’m making the decisions about scope and direction. The AI is doing a lot of the typing, but the shape of the thing is still mine. That was exactly the part I was afraid of losing when I refused to touch these tools in the first place.

It also means I stay current with the final implementation. That makes reviewing Claude’s code much easier, and it means I can jump in and write a piece myself when I want to change something specific.

Context and sub-agents

The other thing Superpowers leans on heavily is sub-agents, and this is where it quietly solves a problem I’d been fighting.

When you ask Claude to go read half the codebase to answer a question, all of that reading lands in your main conversation and eats your context window. A few of those and you’re re-explaining things because the early part of the session has been pushed out. Sub-agents fix this. The exploration happens in a separate agent that does the heavy reading, and only the conclusion comes back to the main thread. My main session stays clean and focused on the actual work. (Yes, the token anxiety from the last post is still very much alive.)

Systematic debugging

One feature I want to call out specifically is systematic debugging.

My old habit, and Claude’s default habit too, is to see a bug and immediately start proposing fixes. The systematic-debugging workflow refuses to let that happen. It forces you to understand the failure first: reproduce it, form a hypothesis, confirm the hypothesis, then fix it. It’s slower for about thirty seconds and faster for everything after that. I’ve stopped losing afternoons to confidently-wrong fixes that addressed a symptom and not the cause.

Plans as a team artifact

Here’s something I didn’t see coming.

The design and plan documents that come out of brainstorming aren’t just useful for me and Claude. They’re a genuinely good way to share context with other people. A plan written in plain markdown, explaining what we’re building, why, and how, is often clearer than what I’d have written by hand for a teammate.

We’ve started experimenting with checking these files into the repo alongside the code, so the reasoning behind a change lives next to the change itself and can be referenced later. It turns out an artifact built to align a human and an AI is also pretty good at aligning two humans.

claude-mem

claude-mem solves the problem of memory across sessions.

The obvious win is efficiency. It builds up a persistent, compressed picture of what I’ve worked on, so when I open a new session I’m not starting cold or re-explaining decisions I already made yesterday. And it does this using far fewer tokens than dumping a whole transcript back into context would. Actually, the “recent context” summary that greets me at the start of a session, the one listing what I did across my projects last week, is claude-mem doing its thing.

But the part I didn’t expect to care about is how it feels. claude-mem keeps track of everything I do, across different projects, over time. It deepens that sense from the last post of working with someone rather than commanding a tool. My collaborator remembers the project, remembers the features we built together, remembers the dead ends we ruled out. It’s a small thing, but it makes the whole relationship feel less transactional.

The other payoff is multi-agent. I don’t only use Claude Code, I also run Codex for a second pass on things. Because the memories live in claude-mem rather than inside one assistant, both can draw on the same history. The context I build up working with one agent isn’t trapped there. Whichever tool I reach for, it knows what we’ve been doing.

rtk

rtk is the smallest of the three, and it does one thing well.

It’s a CLI proxy that sits in front of the commands an AI agent runs, ls, git, grep, find, diff, and a long list of others, and trims the output down before it reaches the model’s context. A raw git log or a noisy build output can be hundreds of lines of stuff the model doesn’t need. rtk strips it to the signal.

The effect is the same as everything else here. More of my context budget spent on the actual problem, less on noise. There’s even an rtk gain command that tells you how many tokens you’ve saved, which is either useful or a way to feed the anxiety, depending on the day.

So…

The thread is the same one from the last post. Superpowers keeps me in the loop, claude-mem gives my collaborator a memory, and rtk keeps the plumbing efficient. None of them try to do all my thinking for me. They make it easier for me to stay in the loop and still be part of the solution, which, for now at least, is exactly where I want to be.