The AI Coding Companion: Vibe Coding Your Bespoke Productivity Universe

Voiceover | 00:00
Are you ready to manage your work and personal world better to live a more fulfilling, productive
life? Then you’ve come to the right place. Welcome to ProductivityCast, the weekly show about
all things personal productivity. Here are your hosts, Ray Sidney Smith and Augusto Pinaud with
Francis Wade and Art Gelwicks.
Ray Sidney-Smith | 00:17
Welcome back, everybody, to ProductivityCast, the weekly show about all things personal
productivity. I’m Ray Sidney Smith.
Augusto Pinaud | 00:24
I’m Augusto Pinaud.
Francis Wade | 00:26
I’m Francis Wade.
Art Gelwicks | 00:28
And I’m Art Gelwicks.
Ray Sidney-Smith | 00:29
Welcome, gentlemen, and welcome to our listeners to this episode. Today, we are going to
continue swimming in the river of AI with our series, The AI-Powered Professional. And today,
specifically, we’re exploring vibe coding. We’re going to discuss really the next major leap in
personal productivity, I think, which is using AI to help you build bespoke tools, workflows, and
interfaces your work actually requires. This is an opportunity now for us to available not only to
professional developers, but to the rest of us, to non-coders and power users who can think
clearly, describe what they want, and iteratively direct AI toward a useful result. Let’s first talk
about what vibe coding is, and then we can get into really just three paradigms, three personas
that I’ve developed for us to bulk together our discussion for today. And so who wants to
describe what they think vibe coding is and or ask any questions about what vibe coding is?
And then we can move on from there.
Francis Wade | 01:35
Nominated.
Art Gelwicks | 01:37
Yeah, you go ahead. You set up.
Ray Sidney-Smith | 01:40
Let’s start with vibe coding having no official definition. There are going to be people who define
vibe coding very differently. Vibe coding, from my perspective, is a way for folks to use some
kind of artificial intelligence to support their software development practice. Now again, this is
one of those cases where software development is a very broad term for what we’re doing here.
Just so that everyone is aware, this term was coined in 2023 by Andrei Karpathy. Andrei
Karpathy is one of the co-founders of OpenAI—he worked on their deep learning and computer
vision—and then later joined Anthropic here in 2026. Be that as it may, he coined the term “vibe
coding”, and the way in which he describes it is as a form of coding where you can “fully give in
to the vibes, embrace exponentials, and forget that the code even exists.”
This is basically a way of not writing code manually. That means it’s a very broad definition and
gives way to a whole bunch of potentiality because now you’re using your cognitive effort and
natural language in order to come up with code. Generative AI models today speak code and
have been trained on so much code that it’s very easy for them to build software in certain ways.
We can use those capabilities to our own best effort.
Art Gelwicks | 03:33
This whole episode, I’m going to be up on my soapbox about certain things because coming
from old-school code writing decades ago, this to me screams of—for lack of a better
term—executive coding. Being able to say: “Just do the thing, and I’ll look at the thing that you
did and tell you how it’s not right, and then you can do it again.”
There’s no structure, no thought process involved, no ramifications or planning. It removes all
the logical friction necessary to provide something substantive and supportive. That’s exactly
what’s desired right now: to eliminate those things and not have to go through that routine and
rigmarole. You want to be able to just say, “Make it.” That’s great if you trust the thing that is
making it to do so in a way that is sustainable, supportive, and reliable.
Ray Sidney-Smith | 04:32
I argue for the exact opposite, which is that this is about applying cognitive effort to solve small
productivity problems in your ecosystem, not attempting to replace professional software
developers. I think it would be very foolish to think this is going to replace a software developer.
This is about enabling software developers, and for non-coders, enabling us to infuse code into
our world where limitations previously existed.
Let me define three personas for today’s discussion:

1. The Designer: These are non-coders. Your main skill is describing what you want with
   precision to define the shape and function of lightweight tools or scripts (e.g., Google
   Apps Script or Excel macros).
2. The Integrator: Power users who understand configuration files (CSVs, JSON, XML) and
   can connect tools together to accelerate custom workflows.
3. The Developer: Professional software engineers using AI companions to refactor, review,
   and generate code while focusing on architecture and quality.
   With that, let’s talk about the non-coder level and what they can achieve with AI.
   Francis Wade | 08:35
   I think the world of the non-coder is the one that’s changed the most. All of a sudden anyone
   can code, which means anyone can imagine solutions. Even if they aren’t fully successful
   initially, getting people to think in terms of software, automation, and coded solutions at a small
   personal level is huge.
   You don’t need to take a college course or pass an exam. The learning curve isn’t as steep
   anymore, so you can start playing around right away to solve daily problems or present small
   incremental answers to clients or managers. That capability feels like having a virtual software
   firm at your fingertips.
   Ray Sidney-Smith | 10:41
   Francis Wade | 10:45
   It’s incredible to have that kind of power at your fingertips. We are in a shift that will change how
   people approach their work.
   Ray Sidney-Smith | 11:13
   I would say we are in a technological evolution. Over the past 20 years, we’ve reached an
   inflection point where natural language interfaces allow us to communicate intent directly to
   computing systems.
   However, non-coders must be cautious. Large language models are probabilistic,
   non-deterministic software. Math and precision require deterministic software (like accounting
   tools). The goal for the non-coder is to learn basic programming principles—inputs, outputs,
   logical branching, and structured formats like Markdown—so you can craft precise instructions
   and think through your workflow logic clearly.
   Art Gelwicks | 15:10
   Here’s the insidious part: this group stands to reap the greatest benefit, but also faces the
   greatest risk. AI tools are marketed with the promise that “you don’t have to know anything, just
   tell it what to do.”
   Take Excel formulas as an example. An AI can generate complex formulas, but if you don’t
   know enough to verify the output, you risk relying on incorrect calculations. You have to treat AI
   like delegating to a human assistant—you need checks and balances to verify that what was
   built actually matches your requirements.
   Augusto Pinaud | 17:31
   It reminds me of when teachers told us in school that we wouldn’t always have a calculator in
   our pocket. It wasn’t about the tool itself, but understanding how to verify the result. We need to
   pay close attention to output quality, whether it’s checking drafted communications or verifying
   references, to ensure AI hasn’t hallucinated plausible-sounding facts.
   Francis Wade | 19:27
   That highlights the need for user judgment. AI gives us incredible technical capability, but lacks
   domain judgment. When non-coders jump straight into building complex systems without
   technical guidance or experienced oversight, they risk creating unmaintainable setups because
   the tool won’t stop them from making poor architectural choices.
   Art Gelwicks | 22:17
   Exactly. Many vibe coders haven’t experienced major system failures or database corruption, so
   they don’t know what a mess looks like. When something breaks, asking the AI to fix it might
   introduce new bugs if there isn’t a structured review and rollback plan.
   Francis Wade | 23:22
   Some people prompt AI to act as its own reviewer by asking it to double-check work against
   guidelines or software standards. Having multi-agent setups or explicit review steps helps add
   that necessary layer of verification.
   Art Gelwicks | 24:37
   Multi-agent setups where tools cross-check each other can help, but you still need oversight to
   ensure the whole pipeline functions reliably without running up unnecessary token costs or
   compounding errors.
   Ray Sidney-Smith | 25:40
   In the Designer persona mode, you should focus on designing lightweight add-ons, scripts, or
   data transformations. For example, exporting data from a task manager like Trello as JSON or
   CSV, and passing it into an LLM with a structured weekly review prompt. The AI reads the
   unstructured data and facilitates the workflow without needing complex software architecture.
   Moving on to the Integrator persona—this is where power users create bespoke internal tools
   or browser extensions. As an example, I created a custom Chrome extension to organize and
   manage my browser extensions. By writing a detailed specification prompt, I generated a
   lightweight, functional extension tailored exactly to my workflow needs.
   Art Gelwicks | 31:58
   That’s an interesting point. Spending time upfront to write an explicit, well-structured prompt is
   essentially doing the requirements gathering that a software developer would perform.
   Ray Sidney-Smith | 32:57
   Exactly. You can prompt the AI to help guide that requirements phase by asking it: “What
   information or constraints are missing from this spec?”
   By defining exact parameters—such as permissions, execution scope, and performance
   requirements—you guide the AI to output clean, efficient code. Upfront effort spent defining
   design specs saves significant friction downstream.
   Francis Wade | 37:04
   I’ve experienced a similar process when designing AI personas for an upcoming conference.
   Managing the strategic design in one conversation while executing specific language or script
   builds in another helped keep the models focused and effective.
   Ray Sidney-Smith | 42:49
   The key is active management. The AI is only as good as the context, constraints, and subject
   matter expertise you feed into it. Start in domains where you possess expertise so you can
   accurately evaluate outputs.
   Francis Wade | 43:49
   Developing that discipline is essential so you know when to trust an output and when to push
   back for verification.
   Ray Sidney-Smith | 45:10
   Maintain a system of master prompts and clear workflows. Treat AI output as a draft that
   requires review. We will continue this discussion on vibe coding and building personal AI coding
   companions in our next episode.
   Thank you, gentlemen, for the conversation, and thanks to our listeners. Join us online at
   productivitycast.net to leave comments or join our community.
   Voiceover | 50:42
   That’s it for this episode of Productivity Cast, the weekly show about all things personal
   productivity, with your hosts, Ray Sidney Smith and Augusto Pinaud, with Francis Wade and Art
   Gelwicks.