Parker Rex
All videos
Parker Rex

This AI Coding App Ships Before Cursor, Windsurf, Copilot

March 14, 2025

Watch on YouTube@parkerrex

Parker Rex dives into Klein, showing how its browser-enabled debugging, plan/act workflow, and logging utilities can streamline AI-assisted coding. He covers practical tips for keeping context under control and steering LLMs toward task-oriented work.

Klein’s core value and what to expect#

  • Built-in browser tools that let Klein open pages, inspect console logs, and “self-fix” code by navigating the app like a human debugger.
  • Focus on keeping context and avoiding LLМ spam by steering the workflow toward task-oriented steps.

Built-in browser debugging and self-fix#

  • Klein can browse, click around, and identify busted vibes in code, then propose fixes directly from within the tool.
  • Console logs and relevant outputs are surfaced to help you see what’s being debugged and what was changed.
  • This is particularly helpful for quickly validating fixes without switching environments.

Logging utility, levels, and tags#

  • A centralized logger with multiple verbosity levels and tags improves visibility across debugging sessions.
  • Use different levels to filter noise and surface only the most relevant information during debugging.
  • This approach makes it easier to reuse and adapt logs across tasks and projects.

Example concept:

  • Implement a simple logger that supports levels (e.g., error, warn, info, debug) and tags, and respects a global verbosity setting.

Code block (conceptual):

const LOG_LEVELS = { error: 0, warn: 1, info: 2, debug: 3 };
let currentVerbosity = LOG_LEVELS.info;

function log(level, tag, message) {
  if (LOG_LEVELS[level] <= currentVerbosity) {
    console.log(`[${tag}] ${level.toUpperCase()}: ${message}`);
  }
}

Context management and avoiding LLМ sprawl#

  • Klein emphasizes tight context windows to prevent overwhelming the model with too much data.
  • The workflow leans toward task-based prompts rather than blasting the LLM with everything at once.
  • You can monitor context window size and adjust as needed to keep responses accurate and actionable.

Plan mode vs Act mode: a practical workflow#

  • Plan mode: use architecture and planning prompts to outline the solution (architecture diagrams, step-by-step plans).
  • Act mode: execute the plan, iterate, and handle results. The switch is seamless and designed to leverage higher-level reasoning before brute-force running.
  • This separation helps maintain quality and reduces unnecessary back-and-forth.

Practical tips Parker uses#

  • Keep a to-do/task list and let the tool steer you toward finishing a task before branching to the next.
  • Leverage the “read me” style outputs to extract actionable instruction sets and tailor them to your project.
  • Parse and repurpose tool outputs into your own instruction sets to speed up future tasks.
  • Be mindful of context length and redact sensitive data (e.g., passwords) when sharing logs or screenshots.

Community, open source, and tradeoffs#

  • Acknowledge the open-source ecosystem around these tools; community contributions can surface powerful capabilities quickly.
  • Parker notes that not all tools dominate in every context, so staying aware of alternatives like Cursor and Windsurf helps you pick what fits your workflow.

Where Parker takes this next#

  • He plans deeper, long-form explorations on topics like plan/act workflows and the nuanced use of logging and context management.
  • See his longer-form content on The Daily uploads channel if you want a deep dive into building an AI services agency and related strategies.

Takeaways you can action today#

  • Try Klein for browser-based debugging and self-fix flows to shorten feedback loops.
  • Introduce a simple logging utility with levels and tags in your projects to improve debugging clarity.
  • Use plan mode to architect solutions and act mode to execute, switching between them as tasks demand.
  • Monitor and manage your context window to keep LLM responses relevant; avoid exposing sensitive data in logs.
Transcript

thinking of reaching out to Klein for sponsorship because I just am going to start talking about them so much my name is Parker Rex I led product and Technology at a startup that did well over $50 million in sales per year until we exited after that I did a bunch of dumb startup ideas but I was using AI the whole time so I feel pretty qualified to be able to talk about this stuff been doing it pretty much non-stop the last three years before that seven years of ux and product management and some coding experience so real quick I just wanted to show this off so I have a thing that was Vibe coded AKA did it wrong and now I'm using Klein to fix things and you'll notice that it's been trying to debug some off stuff and it has browser tools built in like I feel like people just don't know about this so when you're going and you have something that you're running and you want it to try to fix itself it will open up its little browser it'll click around and it'll find oh yeah the vibe code was broken but that's awesome that's like a really big deal and it just kind of goes through it it has the console logs there if there were some and just some other features this is just literally going to be like a one minute video but some other stuff that I find really helpful with it is you have the instructions just as you would expect with like a cursor or a wind surf but there's things in here that are best practices that I feel like other tools aren't implement imp or other people aren't implementing which is to have a loging utility and to have it use that on pretty much everything that it's debugging and then have in that logger utility different levels to it so if you look and we check the logger you'll see that it has levels of verocity and then different tags this is nothing new if you're coding you should be already using this but it makes it even better when you're debugging and then I wanted to just quickly show on clim they released two things that are really really interesting where there's kind of this AI war going on between cursor and wi surf where everyone's like oh my gosh like there's so much better context blah blah blah and then client just like hey someone in our community just made this awesome thing that might be better than all of yours I haven't tested it yet but it's just more of like a a call out that open source is dope and a lot of people are are contributing but what I wanted to show is this so techniques that power users are using so what will apply to all tools is that it really keeps you on top of context windows so it's like you know the task based Paradigm is more helpful because it will push you in the direction of not spamming an llm where you know after a certain point it's going to get bad it recommends after I get to 150 to basically start sweating let see if it fixs itself oh you can see my password I'm going have to remove that let me snap so I remember unexpected error but yeah it's pushing itself in this direction of of basically task oriented work rather than just spamming through it so when it finishes up a task it'll actually push you in the direction not to continue down that route and because again I have this whole thing up here which this is probably the most expensive query I've ever done know why but but I can keep an eye on how large the context window is how much stuff's going in how much stuff's going out and then the plan mode obviously sick because I can go and I can toggle between these two so in plan mode I'm using the all the goodies of of like architecture models or Chain of Thought models that do the architecture for you and then it tees it up for ACT so as soon as I get that context then I click over to act it's like and it starts running it's just fantastic this is what I already mentioned where you can crank up the verbosity it mentions having a read me in every single floor so what I did is I basically just took this and then I parsed out the pieces that are relevant that aren't you know like this is written for somebody that's a human that's reading it but then you just take it and you flip it remove some of the sentences and then put it into your instruction set and you get something that's really really helpful so I'm going to make more videos that are in depth on this especially this right here because I'm really interested in that but just wanted to make a quick video I feel like it doesn't get enough love and I went more in depth on it in my daily upload video on The Daily uploads Channel where it's like long form hourong videos where I basically just am building building public kind of cringe but that's kind of what I'm doing I'm scaling an AI Services Agency to 100K a month and that's so that I can fund my own startup and I talk about the strategy and the things that I'm learning in the field while I'm doing client work because I feel like a lot of content is fluff on YouTube no offense to the people that are just riding the AI hype wave but maybe some offense to you because you're tricking a lot of people you'll be like hey AI runs my whole life and join my school which ironically I have a school I don't charge for it yet but I'm a practitioner I'm in the field I've been in the field for a long time I'm not just like a guru so you can check out that channel too where I go into it in much more depth and that's it for now go check out Klein it's sick and I'll see you in the next video and I got to move over here so I can talk about Klein maybe take the headphones off that's how you make a thumbnail see you