Me, Myself & Claude

How I Leverage AI-Assisted Development to Scale My R-based Data Science Consultancy

Jasmine Daly
Principal Consultant & Founder
Daly Analytics

2025-11-12

Who Am I?

• R Developer & Data Scientist since 2014
• Maintainer of 4 R Packages: {shinyLP}, {ttbbeer}, {avilistr}, {shinyfa}
• Principal Consultant focused on delivering solutions with data science, analytics & AI for nonprofits and mission-driven orgs

Hi everyone, I’m so excited to be here with you all, presenting my talk titled “Me, Myself, and Claude: How I Leverage AI-Assisted Development to Scale My R-based Data Science Consultancy.” I’m Jasmine Daly.

A little bit more about who I am. I’m an R developer and data scientist since about 2014. I’m the maintainer of four R packages, and we’ll go over the last two in the upcoming slides as well.

I’m the Principal Consultant at Daly Analytics, focused on delivering solutions with data science, analytics, and AI. I’m focused on non-profit and mission-driven organizations, handling both the technical and the writing aspects of the consultancy, as well as the sales and marketing aspects.

And just to level set everybody, I’m also the mom of two young kids. I’m an active volunteer in my community. So leveraging AI-assisted development in my practice has been a real game changer, and I’m so excited to share some of my thoughts and concepts about how I use AI-assisted development, specifically Claude and Claude Code.

🔑 Moving Beyond ‘Vibe Coding’

Words matter. Let’s reframe the conversation.

• “Vibe Coding” → Implies reckless, hope-based development
• “Vibe Engineering” → Intentional, accountable AI-assisted development
• “I’m adding a new (agentic) tool to my tool box” → Honest acknowledgment of exploration of this phase of hypergrowth and learning

I came upon this post and this image on LinkedIn, and it was a really great description of a lot of thoughts that have been flying around in my head, really around moving beyond “vibe coding.”

Vibe coding is this funny little term that’s popped up, but it’s really important that we understand that having the proper language helps us to orient the conversation around the work that we’re actually doing, besides some of the fun silly stuff or the mistakes that we’re highlighting.

Vibe coding often implies reckless, hope-based development. Vibe engineering is intentional, accountable AI system development. And what I like to phrase about how I’m doing this is using a lot more words and not just using the small little phrases. I’m adding a new agentic tool to my toolbox. And it’s an honest acknowledgement about the exploration phase of this hyper-growth mode of learning and development, and it really helps reframe and understand and not devalue the work that I do and the work that I’ve done before AI has come into place.

And so I really like for us, as we’re going forward and thinking as you’re seeing the rest of the slides, this is all a new tool in our toolbox, a new way of thinking. It’s not reckless, it’s not hope-based. There’s a lot of deep intention in what I’m building for my clients and what I think a lot of you are also building.

Part 1: Strategic Thinking Unlocked for Greater Community Contributions

The Mental Bandwidth Shift

Before AI

Client Dev

85%

Biz Dev 15%

→

AI Unlocks
15%

With AI

Client Dev

70%

Biz Dev

25%

Community 5%

The mental bandwidth shift of where I was at, probably around this time last year before I was using AI in my code development and my practice. As the principal consultant and developer, I spent an enormous amount of time writing code. Totally what I signed up for, brings me a lot of joy. But it really doesn’t allow me to spend meaningful time doing business development, networking, community aspects, and the real sometimes joyful aspects of this work.

And what I found, looking back from last year to now, with AI integrated into my development process, I get to steal a little bit of that time back and partition it across more emphasis on business development and business operations, which is really important for scaling and growth. And as well as getting back into the community and R/stats community, which really brings me a lot of joy and really helps connect up the “why” of what I do and giving back.

And so it’s been a really interesting shift about having the same amount of time—limited at that—and where it goes towards and what I can now do, now that I leverage AI-assisted code development.

What Mental Bandwidth Buys You

• Better Biz Ops: Improving my systems to be automated and Standardized
• Financial Clarity: Time to understand cash flow, pricing strategy, and profitability
• Relationship Building: Networking, content creation, community engagement
• Creative Space: Room to think about what could be built, not just what must be built

So what that really buys you, to kind of flesh it out: better business operations. I get to spend a lot of time improving my systems, automating things, standardizing things. Coming from being a developer to now being a business owner, efficiency is key where I spend my time. It’s really important, and it’s important to have systems that match up the way I think and the way I want to operate this business. And it really creates a sustainable, thoughtful way that doesn’t also lead to burnout.

Having a lot of time to understand cash flow, pricing strategy, profitability, packaging the services up that I do—it really opened up a lot more time to think about these aspects.

Relationship building. I’ve been able to go to so many more in-person events, virtual events like this one, creating content and posts and community engagement. And it’s been so wonderful to get back out there and meet more people and hear about what they’re doing and the ways in which they’re solving their problems.

Creative space—really the room to think about what could be built and not just what must be built. It’s a really fun time I’m finding to not be just super heads down in what I’m doing, but really think about, well, if I had more time, if I had a couple more hours, what could I build? That could be interesting, that could be fun, that could be silly, that could change how somebody else goes about their work. And so that’s been really exciting to have that creative space back.

The {shinyfa} Package

Client Problem

Months to navigate large Shiny app

↓

Quick Script

Analyze file structure

↓

CRAN Package

Community tool

Before AI: “Nice someday project.”

With AI: Built and released in weeks.

Output:

file_name	type	name
ui.R	input	selectInput
ui.R	output	plotOutput
server.R	reactive	filtered_data
server.R	render	renderPlot
modules/map.R	input	sliderInput
modules/map.R	render	renderLeaflet

Catalog of render functions, reactive functions, inputs, and file relationships across your entire Shiny app.

So I’m going to go over a package that I wrote called ShinyFA, which was released earlier this year. So, I had a client, and it took me a very, very long time to onboard to their application. Grateful for their patience and their guidance, but it was a massive, massive Shiny app. Lots of different files, interconnected functions, the structure of it was challenging. Sometimes you get a shot to your confidence when you’re kind of fighting file systems and folders and just not where you want to be.

And so I thought, “Hey, I’m writing Shiny, and that’s great, but also I can leverage some of my other skills in R. Why don’t I apply that to the overall onboarding and getting to understand what I’m looking at?” And so that worked into a quick little script, a collection of functions where the output on the right-hand side is the file name, and the type, and the name. And having just this level of information in this format was a game changer. It really helped to connect up what I was doing, what changes I was making, how I needed to understand different parts of the code.

And so then that led to the thought process: if it’s helping me now, could it help somebody else? And so I turned it into a package which is published on CRAN, and I really hope all y’all can test drive it and let me know if it also helps improve your development when it comes to navigating large Shiny applications.

And so, yes, before AI, it’s a nice “someday” project, but with AI, I built this and released it in weeks.

The {avilistr} Package

New Hobby

Started birding, found taxonomy in Excel

↓

Wrap Excel

11k+ species, Cornell Lab codes

↓

CRAN Package

In a weekend

Before AI: “Nice to have, not crucial.”

With AI: Published with docs in a weekend.

Output:

Passeriformes: 6,700+ species (over half of all birds!)

The next package that I developed recently this year is called AviListR, and it’s not a super technical lift. Being really forthcoming about that. I have this new hobby this year: birding. It’s fun, it’s relaxing, being in nature, enjoying and being around birds.

And so I found this newly released taxonomy sheet in Excel. And it’s great, right? Excel—we can all work with Excel for the most part. But I wanted to have this opportunity to think, “Well, can I make an R package?” That’s silly—like, an R package with bird data. That is not a large technical lift, but maybe it could be more accessible for folks. And so it’s a really large Excel sheet, I wrapped it up, and it has some really interesting aspects.

I built this in a weekend. For me personally, I was never moving at this speed for a CRAN package. The two previous packages I published, ShinyLP and TTBBeer, were published in 2016. AviListR, which was published in June on CRAN this year, that is eight years and nine months between my first set of R packages and this new one.

I feel this sense of hyper-productivity, excitement, and just having the aspect of integrating AI in development for packages and community-based tools has been really transformative. And I know there’s a lot of things that were going on within the last eight years and nine months, but I really want to underscore what I’ve been able to do just this year in giving back to the community and thinking grander about how do I contribute open source code.

And so, yes, it’s not even a very crucial package in my opinion, but it is nice to have. It is fun. And it now has really lovely examples about how you integrate this data with some of the existing packages like rebird. And so this is my little passion project and my gift to the R community.

🔑 The Creative Multiplier

AI assistance doesn’t just save time on what you were already doing. It unlocks what you weren’t previously able to complete before!

• Automated Biz Dev and Biz Ops systems
• Writing better unit tests, documentation, package branding, hex stickers
• Community engagement and knowledge sharing, networking

So really holding on to this: AI-assisted development is the creative multiplier. It doesn’t save me time on what I was already doing. But it really unlocks what I wasn’t able to do—what I wasn’t able to complete, what I wasn’t able to prioritize, simply because when you’re running a solo consultancy and being really strategic about your time and protecting your mental health, very often sometimes you don’t get to do the really fun things.

And so now that I’ve been able to leverage it in my work processes, I’ve been able to automate all those biz dev and biz ops systems. I can write much better unit tests and documentation and package branding and hex stickers. Shout out to ChatGPT—I find it really wonderful to collaborate on making those really engaging hex stickers. And then just having more time for community engagement and knowledge sharing and networking.

That for me is a really important aspect of the work I do and how I show up. And I’m so glad to have been leveraging AI-assisted development and Claude Code in my process. Because I feel like it’s not just me benefiting, which is great right now. It is everybody. It is all of our learning paths. It is all of our creative endeavors. And to be able to provide this for my clients and also provide it for our community, which I love so much, has felt really, really meaningful.

{ttbbeer} my first package was released July 2016 and {shinyLP} my second package was first released on September 2016. The next CRAN package i released was {avilistr} in June 2025. That’s 8 years 9 months (September 16, 2016 – June 17, 2025)

Part 2: Saying Yes to Bigger & Funner Problems

Optimizing for Professional Joy

This is an aspect of how I think about work, whether I’m working for myself, for my consultancy, or doing other types of volunteer work or other types of work in the community.

And so it really sits around the client work—working with values-aligned clients, working with non-profits and mission-driven organizations. These are organizations that are doing really important boots-on-the-ground, highly interpersonal work in this world. And I love being able to be in service to them.

I also really love opportunities to grow technically, work on problems that I may not have the full answer for but I have excitement around, exploring new approaches. This work keeps things interesting and mentally engaging.

Another aspect is community professional development. I love going to networking events, I love going to conferences, and it’s really important for me in that aspect.

And as well as intentional rest. All of this does not work unless I have a really good sense of when is work and when is family and friends time, because all of this is in service for that true goal.

And for me, when all of those things hit, that is professional joy. And that is what I optimize for in my consulting practice. And it really drives me in the decisions I make.

The Old Filter

Before AI-assisted development, my filter for new projects was:

“Can I build this in a reasonable timeframe with my current skillset given my existing client workload?”

So my old filter before AI-assisted development: when coming to a potential new project or a prospect or a client, “Can I build this in a reasonable timeframe with my current skillset given my existing client workload?” That’s what I told myself.

The New Filter

With AI-assisted development, my new filter has become:

“I can certainly build it! How can I thoughtfully spend time deeply understanding the desired outcome & client constraints?”

Now, with leveraging Claude and Claude Code and agentic development, my first thought is, “I can certainly build it. But how can I thoughtfully spend time deeply understanding the desired outcome, the client constraints, and the overall delivery?”

Claude Code as My Coding Partner

The Mental Model

• I’m the manager: Setting direction, making architectural decisions
• Claude is my senior developer partner: Implementing solutions, handling boilerplate
• Human in/on the loop: Active technical leadership stays with me

Tips & Tricks for Managing the Partnership

• Be specific with requirements
• Always review generated code
• Use incremental commits
• Feed documentation links
• Ask for multiple options
• Run Claude in parallel across related repos

With Claude’s help I amplify my capacity to take on bigger, more interesting problems.

I see Claude really as my coding partner, and the mental model around it is: I’m the manager. I set direction. I make the architectural decisions. Claude is my senior—sometimes my junior developer, depending on the context that you give it—a partner in that endeavor. Implementing solutions, handling boilerplate code, starting out any of the scaffolding for packages or Shiny applications.

And really seeing Claude Code as my coding partner means that I am both the human in the process and on the loop. So in and on the loop—active technical leadership stays with me. I’m integrated at various different intersections in the entire process, and that is table stakes for me.

And so ways in which I help manage this partnership: I must be really specific with requirements and instructions for Claude. I always review the generated code. Claude has this wonderful way where you can, when you set up your context files, have it just writing code and committing to your GitHub repository and not being a part of that discussion and that conversation. That work pattern, for me, I always need to be in the mix for reviewing code.

And related to that, I always use really incremental commits. I love how Claude can write really detailed commit messages. I love being able to possibly unwind things back and forth, and so that’s super important for me.

And feeding it documentation links has been really helpful to help keep Claude focused and giving it accurate information.

I’m also asking for multiple options. I find really using Claude, in addition to coding, as a brainstorming partner to get me started on a lot of the work that I do.

And then I love to run Claude sometimes, depending on the project, in parallel across related repos. This amplifies my impact to take on bigger and more interesting problems.

Client Example: Using GitHub Actions For Data Orchestration

Data Orchestration Repo

GitHub Actions (API, webhooks, Railway)

↓

Load Data in Supabase

RLS rules per client_id for data isolation

↓

Client Dashboards (Quarto)

50+ dashboards on Netlify, each pulls only their data

↓

Template System

Shared components, feature toggles, GH Action PRs for updates

Right-sized enterprise-grade infrastructure built in a few weeks with Claude Code’s help.

A client example that I really love: they needed to scale a data orchestration system for 50+ client dashboards. After understanding their requirements and technical capabilities, GitHub Actions made a lot of sense. I worked with Claude to build this system—APIs, webhooks, Railway for the VM—really learning as I went.

The architecture: loading data into Supabase with row-level security rules for client isolation—I had a real hand on the wheel there for the data security and governance aspect. Then all those client dashboards written in Quarto, each pulling only their data.

The client needed a repeatable ETL process, secure data layers for persistence, and a shared component template system where they could feature toggle different dashboard aspects on and off. So each client gets their own dashboard, but they all originate from the same template with shared components that can be updated centrally.

I developed this system primarily with Claude Code—lots of back and forth, lots of learning on my part about writing better GitHub Actions and how to have Actions create PRs in other repositories. This is right-sized enterprise-grade infrastructure built in just a few weeks with Claude’s help. This would not have been possible before.

Client Project Workflow

1. Understand Requirements (Human) → Define the problem and technical approach
2. Generate Implementation (Claude Code) → Write the initial code
3. Review & Course-Correct (Human) → Read every line, check for issues
4. Iterate Based on Feedback (Claude Code) → Fix problems, refine solution
5. Final QA & Delivery (Human) → Test thoroughly and ship to client

Human decisions at the start, middle and end. AI execution sprinkled in the middle.

So my usual client project workflow is:

1. Understanding the requirements - defining the problem, the technical approach
2. Generating some of the implementation - Claude writes the initial code
3. Review and course correcting - reviewing every line, checking for issues
4. Iterating based on that feedback - fixing problems, refining the solution
5. Final QA - testing thoroughly and then shipping to the client

It’s human decisions at the start, the middle, and end, and AI execution sprinkled in the middle.

🔑 Ruthlessly Solutions-Focused

AI-assisted development has made me even more ruthlessly solutions-focused for my clients.

• I’m not thinking about syntax—I’m thinking about client outcomes
• I’m not debugging YAML—I’m solving business problems
• My mental energy goes to what matters: understanding client needs and delivering value

The technical details still matter—but they don’t consume my cognitive load anymore.

So I am ruthlessly solutions-focused at this moment, thanks to Claude Code. It’s made me be even more a better partner for my clients, I believe.

I’m not thinking about syntax. I get to think about their outcomes. I’m not thinking about debugging YAML for GitHub Actions. I get to solve business problems alongside the stakeholders. My mental energy goes to what matters—understanding client needs and delivering that value.

The technical details still matter. I’m still very entrenched in writing code nearly every day, but they don’t consume my cognitive load anymore. I am not just a person on the computer. I get to communicate and talk and have these really important interpersonal relationships with my clients.

Part 3: Building Up Your R Intuition Through Volume

Learning Through Volume

Traditional Learning

A few hours for a few appraches

AI-Assisted Learning

A few minutes for many approaches

Fast iteration builds intuition faster than slow deliberation.

I like to think of learning as highways. Traditional learning is a smaller, focused highway—a few hours, you see a few approaches. You’re reading docs, checking Stack Overflow. It can be slow. Meaningfully slow sometimes.

AI-assisted learning? That’s seeing many approaches in a very short time. A lot more throughput through that highway system. But you have to be mindful—some things are going to go awry. You need to distill what’s the core of what you need and what’s just wrong.

Mistake #1: Package Confusion

Claude doesn’t read the docs—just like you when you’re rushing

• What happened: Asked Claude about {shinyfa} → it hallucinated “Font Awesome icons”
• The real error: Like installing a package and guessing without reading vignettes/examples
• It’s not just new packages: Even with {ggplot2}, Claude confuses function arguments
• Why this happens: You’re experienced, busy, assume things should come easy, so you circumvent
• The lesson: This forces you to actually read documentation and verify behavior
• Fix strategy: Feed Claude the documentation URL, then it self-corrects

This mistake makes you better at evaluating package quality and knowing where good docs live.

I’m going to share some mistakes that happened even building this talk with Claude Code.

Claude doesn’t always read the docs—just like humans rushing through things. When I had Claude structure the initial slides, it completely hallucinated what the shinyfa package does. It assumed “FA” meant Font Awesome icons. It didn’t even install the package—just guessed about vignettes and examples.

This happens even with established packages like ggplot2. Experience breeds assumptions, and you skip verification.

The lesson: feed Claude documentation URLs so it can self-correct. This actually makes you better at evaluating where good docs live.

Mistake #2: Data Structure Assumptions

Claude assumes instead of checking—sound familiar?

• What Claude does: Assumes column names, data types, relationships
• Example: Used avilist_global (doesn’t exist) instead of avilist_2025
• Example: Assumed lowercase order_name when actual column is Order
• Example: Infers connections between columns instead of asking for lookup tables
• Why this happens: Experience breeds assumptions; you skip verification when busy
• The lesson: Use mcptools to give Claude actual R session context
• Fix strategy: names(data), str(data), or let mcptools read your environment

This mistake teaches you proper data inspection habits—skills you use constantly.

Claude loves to infer from column names and data types. But sometimes it’s just wrong.

Building the AviListR slide, Claude called the dataset avilist_global—which doesn’t exist. It’s avilist_2025. It also assumed snake_case column names when they actually match the original Excel file’s title case.

Experience breeds assumptions. You skip verification. “Well, snake_case is commonplace, right?”

The lesson: leverage MCP tools to give Claude actual access to your R session. If packages are loaded, if objects exist, let MCP integrate and read your environment.

This mistake teaches proper data inspection habits—skills we should all be using constantly as data practitioners.

Mistake #3: Over-Abstraction

Why is Claude (or you) always doing it the hard way?

• Claude’s tendency: Always reaches for complex solutions first
• Example: Tried 3 different Sankey packages with hallucinated functions
• Example: Creates nested functions when a simple dplyr chain works
• Why this happens: You want to show expertise, build “proper” systems, over-engineer
• The lesson: Simple is better. Ask “Why are you doing it the hard way?”
• Fix strategy: Ask for multiple options, choose the simplest that works

This mistake teaches you that simplicity is a skill—recognizing unnecessary complexity.

This is my biggest gripe with Claude. It always reaches for the complex solution first. Always trying to be overly clever.

Building this slide deck, Claude tried creating an intricate Sankey diagram and just started making up functions. Nested functions everywhere when a simple dplyr chain would work.

I think it’s pulling from examples where people are showing expertise, building proper systems—which sometimes leads to over-engineering. Missing the focus: simple is often better.

So I ask Claude a very direct question: “Why are you doing it this way?” Often, because I’ve spent so much time without AI, I do know the solution. I know the exact packages I want to use.

I ask why, feed it different options, and select the simplest way—the way I would actually do it. Recognizing unnecessary complexity has been key to guiding this partnership.

Building Better Prompts from Mistakes

Each mistake pattern teaches you how to redirect Claude (and yourself)

• Package confusion → Feed documentation links upfront
• Data assumptions → Use mcptools, provide str(data) output, give lookup tables
• Over-abstraction → Ask for multiple approaches, request “simplest solution”

Learning at Scale: You encounter these patterns 10x/day instead of 1x/week. Each failure refines your prompting skills AND your R intuition.

You’re not becoming dependent—you’re becoming a better technical leader who knows how to redirect when things go wrong.

Here’s some tips for building better prompts from these mistakes:

• Package confusion → Feed documentation links upfront
• Data assumptions → Use MCP tools, give them lookup tables if columns are coded in other ways
• Over-abstraction → Ask for multiple approaches, request sometimes the simplest solution to start

Learning at scale: you’re going to encounter these patterns so much more than traditional learning. And each failure, I think, helps refine your prompting skills and your R intuition so that you don’t make those mistakes and so you can help Claude make those mistakes less.

And I think you become a better technical leader, and you know how to redirect when things go wrong and leverage what you’re learning from Claude to bring to the real world as you’re interacting with folks to create better software.

🔑 The Art of the Practice

Skill development is a systematic approach—research, analysis, consistent action

• Not about the final product alone → It’s about the process of getting there
• Research & Analysis → Understanding patterns, learning from mistakes, building intuition
• Consistent Action → High-volume iteration, failing fast, course-correcting
• Achieving Mastery → Through deliberate practice and pattern recognition
• Personal Growth → Overcoming challenges, maintaining motivation, staying disciplined

AI-assisted development accelerates the practice by increasing volume—but the discipline of learning is still yours.

I really like to think of this all as the skill development, and that can be phrased as the “art of the practice”.

It’s not necessarily about the final product alone. It’s the process of getting there. Understanding patterns, learning from mistakes, building that intuition up. High-volume iteration and failing fast and course-correcting. Deliberate practice and pattern recognition. And the personal satisfaction and growth from overcoming those challenges and maintaining motivation and staying disciplined.

I really think AI-assisted development accelerates this practice by increasing volume. But the discipline of learning is still your own.

Conclusion: Human In/On the Loop

Your AI Values Exercise

Before your next AI-assisted project, write down:

1. What I will always do myself
2. What AI can help me with
3. What I need to understand deeply
4. What I’m comfortable delegating

So in conclusion, I really want everybody to hopefully take away this thought pattern of learning for yourself, and then also writing down your values when it comes to interacting with AI in your work, for fun. And before your next AI-assisted project, please go through this exercise. Write it down:

• What will I always do myself?
• What AI can help me with?
• What I need to understand deeply?
• And what I’m comfortable delegating?

Final Thoughts

“When AI frees up your mental bandwidth, you don’t get lazy—you get strategic.”

• Build with intention.

• Stay curious about failures.

• Stay strategic about solutions.

• Stay joyful in the work.

So some final thoughts. I believe when AI frees up your mental bandwidth, you don’t get lazy. You get strategic.

Always building with intention. Staying curious about those failures. Staying strategic about solutions. And staying joyful in the work, because if we’re not all having fun doing this, I don’t think any of this is worth it.

So please always remember: have fun, have joy. Take that intentional rest. Be really mindful about what you’re building, who you’re building it for, and ensure that you’re not causing any more harm through your actions and the software that you’re making.

Thank You

Questions?

Contact: jasmine@dalyanalytics.com

GitHub: @jasdumas

Website: dalyanalytics.com

Thank you all for your time. I’m happy to take questions about AI-assisted development, R workflows, or anything else we’ve covered today.