My $20 AI Subscription is a Personal Software Factory

I had 36 podcasts in my backlog. 54 hours of audio I’d never listen to.

My first instinct: build a tool to consume all 36. Summarize everything, extract key points, process the entire backlog. I used Claude Code ($20/month) to build it in a weekend—Payload CMS, database, text processing, the works.

Four hours into tweaking the UI, I realized I was optimizing the wrong thing.

Think Slow, Act Fast

I use a simple mental model for projects: Think Slow (planning phase) and Act Fast (delivery phase).

Planning phase: Cheap iteration. Test ideas, break things, learn fast. Cost of failure is low—just time.

Delivery phase: Expensive execution. Ship the validated plan with speed and precision. Cost of failure is high—delays, budget overruns, startup death.

AI makes the planning phase faster. I can iterate on ideas in hours instead of weeks. But here’s the trap: it’s so easy to build something that feels like delivery when you’re still planning.

I was polishing a UI for a tool I hadn’t validated yet.

The Problem Evolution

I started with Option A: “Consume all 36 podcasts via summaries.”

While building, I shifted to Option C: “Extract insights without listening to anything.”

What I actually needed was Option B: “Decide which 5-10 podcasts deserve my full attention.”

The tool couldn’t tell me what I needed. AI summaries are generic. They don’t know what I’m working on this week, what problems I’m stuck on, or which ideas will connect to my current experiments.

I needed triage, not consumption. Surface what deserves full attention when timing is right.

What I Use It For Now

The tool is live in the Podcast Section of my blog

I use it to:

• Craft essays (find material worth exploring)
• Decide which videos to watch fully
• Archive podcasts that might matter later (not now)

A podcast irrelevant today might be exactly what I need next month. The tool lets me defer the decision without losing the option.

The Expertise I Needed

Building this required domain expertise, not technical expertise.

I defined the data flow—how podcast content should be organized, what metadata matters, how summaries should be structured. Without this, the database would corrupt with every change.

Claude Code handled the implementation: Payload CMS integration, database schema, API calls, text processing. I couldn’t verify the technical implementation. I validated the logic.

This connects to my AI amplification thesis: You need expertise to define the problem and workflow. AI handles the implementation. But you can’t extract value if you don’t know what you’re looking for.

What I’m Watching For

I don’t know yet if this tool becomes digital clutter. Will I still use it in 3 months? Will the maintenance burden eat the time I saved?

The hard constraint of Claude Code’s starter plan helps. After one wasteful iteration (the 4-hour UI trap), I’m forced to think before spending another slot. Constraints force planning.

Here’s what I learned: AI blurs the line between planning and execution. It’s so easy to build that you forget you’re still planning. The 4-hour UI trap happened because I treated the tool like it was in delivery phase (Act Fast—polish the UI) when I was still in planning phase (Think Slow—figure out what I actually need).

The tool exists at podcast.kiranbrahma.com. Test it yourself. Let me know if you hit the same trap.