Show HN: PMX – A context manager for CLI AI Agents

3 months ago 5

Last Tuesday, 2:47 AM. I'm deep in a gnarly bug in our payment processing system. Claude is helping me trace through the async flow when suddenly I need to context switch - critical security PR just came in for one of my open source projects.

I open the PR, ask Claude to review it, and it starts suggesting we add tokio::time::timeout to handle the payment timeout issue.

Wait, what?

Oh right. Claude still thinks we're debugging payments.

Sighs. Opens ~/.claude/CLAUDE.md for the 50th time this week.

That night was my breaking point. I actually started keeping a tally. Between my day job, three active open source projects, and the random midnight "hey can you review this?" messages, I was editing that damn config file 12-15 times a day.

You know that feeling when you're in flow, crushing code, and then BAM - you have to stop and explain to your tools what you're doing? It's like having to re-introduce yourself to your IDE every time you open a new file.

The math was stupid:

15 context switches × 3-5 minutes each = 45-75 minutes daily
That's 5+ hours every week
260+ hours a year
Just... editing... config... files

I've built tools for smaller annoyances than this.

So there I was, supposedly fixing that payment bug, but instead I'm in my terminal typing cargo new pmx. Because apparently that's what we do at 3 AM - start new projects instead of sleeping.

The idea was dead simple. What if switching context was just:

pmx set-claude-profile payments-debug

Done. Three seconds. No file editing. No copy-pasting. No forgetting what formatting Claude prefers.

Here's the thing nobody tells you about building developer tools - you don't realize how broken your workflow is until you fix it.

First day using PMX: saved 10 minutes. First week: saved 2 hours. First month: I'd forgotten what it felt like to manually edit configs.

But the real win? I stopped losing my train of thought. That payment bug? Fixed it in 20 minutes after I could properly context switch. The security PR? Reviewed with actual security-focused prompts, not payment processing context.

Real Profiles from My Real Life

I'm not gonna show you some pristine, organized folder structure. Here's what my ~/.config/pmx/repo/ actually looks like:

├── GET-THIS-SHIT-DONE.md # When I need Claude to stop being chatty ├── debug-why-the-fuck.md # 2 AM debugging sessions ├── rust/ │ ├── unsafe-cowboy.md # For when I need to write cursed code │ ├── please-work.md # My default Rust profile (yes, really) │ └── macro-wizard.md # proc macro hell navigator ├── reviews/ │ ├── security-paranoid.md # Every input is evil │ └── dont-be-an-asshole.md # For reviewing junior devs' PRs └── work/ ├── payments-prod-ohno.md # Production incidents └── meetings-help-me.md # For writing docs during meetings

Yeah, I name my profiles like a normal person at 3 AM. Sue me.

The One That Started It All

That payment bug? Here's the actual profile I wrote that night (payments-prod-ohno.md):

We are debugging PRODUCTION payment issues. This is NOT a drill. Context: - Rust microservice handling real money - Using sqlx + postgres, NO ORMs - Every error MUST be logged with context - If you suggest unwrap() I will cry Current issue: async timeouts in payment flow Focus on: transaction isolation, timeout handling, rollback safety DO NOT SUGGEST: - Refactoring (we're debugging, not rewriting) - Adding new dependencies - "Have you tried rewriting it in X?" Just help me find the bug. Please.

My Favorite Hack: Profile Layering

Discovered this by accident. Was reviewing a PR at 1 AM, needed both Rust knowledge AND security paranoia:

pmx set-claude-profile rust/please-work pmx append-claude-profile reviews/security-paranoid pmx append-claude-profile GET-THIS-SHIT-DONE

Claude went from "Here's a comprehensive analysis of..." to "Line 47: SQL injection. Fix: use prepared statement. Next issue..."

Chef's kiss

The Unexpected Side Effect

Here's what I didn't expect: PMX made me a better developer.

Not because of some productivity guru bullshit. But because writing these profiles forced me to think about what I actually need in each context.

Like, what makes a good code review? I had to write it down. What are my Rust conventions? Had to articulate them. What should Claude focus on during debugging? Had to decide.

It's like rubber duck debugging, but for your entire workflow.

OK But Does It Actually Save 30 Minutes?

Last week I kept actual stats:

Monday: 18 context switches
Tuesday: 22 (PR review day hell)
Wednesday: 14
Thursday: 19
Friday: 11 (I was tired)

Total: 84 context switches × ~4 minutes saved = 336 minutes = 5.6 hours

So yeah, I lowballed it with "30 minutes a day." It's more like 30 minutes before lunch.

Look, I could write another 1000 words about how PMX uses Rust's type system for safety, or how it respects Unix philosophy, or whatever. But honestly?

Just install it. Make one profile called help-me.md with whatever you wish Claude knew about your current project. Use it for a day.

If it doesn't click, uninstall it. I won't be offended.

But when you find yourself at 3 AM, about to edit ~/.claude/CLAUDE.md for the 15th time, remember this post.

Your future sleep-deprived self will thank you.

Built at 3 AM by someone who should have been sleeping. Find me and PMX at github.com/NishantJoshi00/pmx

Read Entire Article