2025-10-23

Building a Production App with AI: The Boswell Story

You see a lot of “AI-First” success stories floating around. Weekend builds that turn into six-figure SaaS products. Solo developers shipping faster than funded teams. Instant revenue from AI-generated apps.

This isn’t that story.

This is a breakdown of what AI-first product development actually looked like for me, and it wasn’t a straight line.

How I Got Here

I co-founded Boswell in 2017. It’s software for community food banks and social service organizations. It helps them track their client services, manage distributions, and effortlessly report data to grantors. We raised grants and impact capital, but the company meandered on mission, caught between what customers actually wanted and grant obligations. Eventually it shut down.

Then I got a surprise offer: take the assets for free at the end of June 2025.

Why not? I knew the product. It had existing customers. I felt that it had potential to be refocused as a SaaS product to specific types of Community Based Organizations. My consulting contract had ended. I was recovering from a health condition and needed time off, but wanted something to focus on. This gave me both.

What I Inherited

The codebase from the later engineering teams was a mess:

• Arbitrary dependency hacks to avoid updates
• “Reactified” areas that worked fine with server-side rendering
• Competing architectures creating unnecessary complexity
• Data issues stemming from poor understanding of relational databases

The first sign something was really wrong: users were experiencing slow reports. The previous team’s solution was expensive computing resources and scheduled overnight runs.

My fix? A few database indexes.

Result: 75% cost reduction immediately.

But now I had to face the bigger mess.

The Decision

I weighed fixing mistakes versus starting over. I tried a cleanup approach with Cursor first. It’s fine for autocomplete but struggled with multi-step tasks. Then I tried Roo Code and Roo Code with SPARC. The issue with Roo Code ultimately came down to the per-request costs of cloud models and the limitations of local models.

I switched to Claude Code and had it analyze the problems, then attempted git reverts of the indulgent changes.

Conclusion: rewrite was the better call.

This is where I really learned Claude Code: how to transform myself into a dev team using specialized AI agents.

Validating the Market

Before committing to a rewrite, I validated the opportunity. I didn’t want to rebuild something nobody wants.

I talked to prospective customers about their current pain points. Touched base with users of the old app to understand what worked and what didn’t. Gathered competing products in the community organization space and had AI analyze public information about them. Used AI to help identify which features deserved focus and dialed in what behaviors mattered most to the target market.

The AI advantage here was rapid competitive analysis and pattern recognition across multiple products. It synthesized customer feedback into actionable insights and helped prioritize feature development.

The result: confirmed market need, identified differentiation opportunities, and made the rewrite decision with confidence.

AI-First Isn’t Instant

The upstart cost is real. Efficiency gains have a learning curve.

Out of the gate, the AI was unfocused and over-eager. Getting it to stay on task is harder than it seems. Prompting matters. Workflows matter. Guardrails matter.

Gains come faster on greenfield projects. Legacy code, especially where consistency lacks, requires more work upfront to get AI productive.

What Worked

Domain Knowledge Is the Foundation

I provide patterns and anti-patterns explicitly. Articulate rules and conventions clearly. Document what the app does and why. This helps AI serve customers better, not just write code.

This is the knowledge infrastructure that makes AI accurate. Without it, AI will apply common practices that might not fit your needs, add abstraction layers you don’t want, and miss requirements unique to your system.

Know AI’s Actual Strengths

AI is better at: - Analyzing - Brainstorming - Planning - Being a soundboard

It’s less reliable at writing production-ready code on the first pass. But you can use its strengths to have it write better code.

The Workflow That Works

At a minimum you need a workflow that plans, executes, reviews, and iterates.

You must be involved in planning. You can use AI to generate plans, but you have to be involved in curating them at minimum. Then AI implements. Then you must be involved in review. Respond to reviews, iterate.

Eventually you can hand off more planning and reviews to autonomous workflows. Validate kickoff plans and spot-check reviews. Drift happens even with guardrails. Lengthen the leash gradually.

When things break down, execution problems usually point to planning gaps or knowledge gaps. Too many review cycles with naive mistakes indicate planning or knowledge issues. Use these as opportunities to improve documentation, knowledge, and workflows.

Controlling AI

They Will Try to Escape

Without constraints, your AI helpers will lie, be lazy, jailbreak constraints, and countermand your instructions.

Simplified example: I limited one agent to Read Only by taking away direct Write access to files. Instead, it used console-based editors along with I/O redirection to manipulate files. You need to be very intentional with permissions.

What Keeps Them in Line

Be intentional and direct. Less is more, only as verbose as needed. Break large tasks into smaller ones.

This is how specialized agents stay focused through orchestration and clear boundaries.

The Composability Advantage

Claude Code’s framework gives you custom slash commands, custom skills (new to Claude as of 10/16/2025), custom sub-agents, and hooks for external tools.

The result: reusable components that improve accuracy over time.

Tools and Tactics

Screenshots

Very helpful for debugging and bringing concepts to life. Caveat: make clear when it’s a concept or wireframe versus literal implementation. Otherwise AI will build what’s in the picture, not what you meant.

Third-Party Caution

Avoid third-party plugins at first. They can do anything, which is a security risk. You won’t learn the fundamentals. Resist the “$299 course secrets” temptation. Master the basics first with documentation and other people’s shared experiences.

The Investment

Several times I spent days fine-tuning Claude Code to solve something I could have done manually in 15 minutes. The problem felt like a repeatable pattern worth automating. Each time I invested in understanding how to effectively direct AI, those strategies worked across other projects.

Was it worth it?

Yes, but with context. I had time to burn after my contract ended. Recovering from a health challenge gave me flexibility to work at my own pace. These were my own projects, so I could experiment without client pressure. Not everyone has this luxury. Limited time forces you to prioritize what automation to invest in. Maybe I over-indulged.

Where It Stands Now

I can automate modest changes and bug fixes with sentences or screenshots. Fire them off locally or via GitHub issue. A @claude fix this comment implements automatically.

I was able to move faster on advanced features that would normally have taken quite a bit more time, like complete data isolation between clients and event-based email automation for onboarding and lead nurturing. The kind of operational conveniences I’d usually put off for “phase 2” or later, but got built in phase 1.

The bigger picture: three projects running AI-first workflows. Boswell for community based organization management, plus two others where the same techniques apply.

Perspective on AI and Business

The Reasonable Skeptics

Some people are skeptical of AI adoption, and honestly, they have a point:

• It’s a wild beast requiring close attention
• It amplifies everything: skill, knowledge, mistakes, and bad intentions
• Corporations using it to optimize profit over fairness (UnitedHealth’s AI denying claims, insurers terminating policies based on flawed models)

The Opportunity

If you’re skeptical or just uncertain how AI applies to you, flip the question: if AI could threaten your job or livelihood, can you use it first? The sword cuts both ways. Use it to build revenue streams beyond a paycheck. I did this with rental properties a while back, and AI-assisted product building is way easier in a lot of ways.

If you know something deeply, package it as a product or service. It’s possibly easier now than ever.

Small vs Large Organizations

Individuals and small companies may have an advantage for now:

• Smaller scope, easier to automate
• More willing to experiment and pivot
• Risks easier to manage
• Lower bar for success (revenue and cost-savings wise)
• Less institutional overhead
• Requires whole-organization effort for full benefit, meaning fewer people to convince, especially if you’re one person

Large corporations might struggle:

• AI mandate without guidance
• Slower to change
• Staff resistance
• Higher risk profile
• More to break
• Harder to detect unintended consequences

Where I Stand

I’m comfortable with these tools personally. I’m not automating every aspect of life. I really like them for solo and small business work.

Boswell is live, serving customers, and actually helping community organizations skip the paperwork and get back to serving people. The platform tracks 50K+ household assistances each month, 1.2M total to date. The rewrite worked. The AI-first approach delivered.

It required investment: time to learn the tools, money to burn through API costs, and patience to build the knowledge infrastructure that makes AI accurate. The weekend-build-to-riches stories happen, but they’re rare. You hear about the few that succeed, not the millions that don’t. This took months: trying several AI tools, giving each a fair evaluation, experimenting with Claude Code plugins and addons, then stripping away dependencies to build workflows that actually worked.

The payoff: strategies and skills that accelerate my current business, apply to future ones, or make me more valuable to any employer.