Nobody Laid Out The Five Kinds Of Software You Can Make. So I Did. (Video)
Source: YouTube
Ingested: 2026-08-26
Executive Summary
In this comprehensive architectural guide, nate-b-jones demystifies the software development lifecycle for non-technical builders in the era of AI coding agents. Nate introduces a structured taxonomy of Five Software Shapes—ranging from single-file local scripts to hardware-integrated edge systems—and explains how non-developers can use modern coding harnesses (claude-code, codex, Lovable, Replit, Bolt) to build custom personal-software.
Nate emphasizes that successful software creation no longer requires memorizing technical syntax; instead, it demands clarity of intent, good context-minimization and data architecture, and treating AI models as collaborative building partners that must explicitly reveal and document technical trade-offs.
Key Tactical & Strategic Takeaways
1. The Era of Personal Software
- Commercial mass-market software is built for broad distribution and average workflows, often failing to address idiosyncratic, hyper-specific individual or household problems (e.g., custom appliance maintenance tracking, custom ferry radio tracking via AIS).
- Coding agents enable ordinary individuals to build personal-software tailored specifically to their immediate physical and digital environments.
2. The Five Software Shapes Taxonomy
Nate categorizes all software builds into five primary shapes to guide architecture and tooling decisions:
- Local Tool: Runs locally on a single machine, interacts with local files or SQLite. Zero server management, zero authentication overhead (e.g., file organizer, private research tool).
- Web App: The recommended default for personal software. Runs in a browser, accessible across mobile and desktop without App Store friction, using hosted backends like Supabase or Lovable Cloud.
- Native Phone App: Reserved strictly for tasks requiring deep hardware APIs (background location, NFC, push notifications, Bluetooth) via frameworks like Expo; involves App Store approval friction.
- Background Service / Scheduled Job: Headless process executing on a schedule or webhook trigger (e.g., polling public feeds, automated daily summaries).
- Hardware Project: Software running on edge microcontrollers or single-board computers (Raspberry Pi, ESP32, Home Assistant) to interface directly with physical sensors (e.g., radio receivers, environmental monitors).
3. The 4-File Project Steering Framework
To prevent coding agents from making opaque, expensive, or destructive technical assumptions, non-technical builders should maintain four markdown files in the project root:
project.md: Plain English problem definition, user personas, current vs. desired state, deployment target, and privacy boundaries.decisions.md: Append-only log of architectural choices (database selection, auth model, hosting) documenting options considered and rationale.scenarios.md: Concrete end-to-end user scenarios and edge cases acting as functional test criteria.claude.md/agents.md: Operational instructions directing the coding harness (claude-code or codex) to explain technical trade-offs in plain English, ask before making destructive changes, and protect secret keys.
4. Builder Ecosystem & Stack Decoupling
- Builders should decouple the Model (reasoning capability, e.g. Claude, GLM, GPT), the Harness (execution environment, e.g., Claude Code, Codex), and the Host/Backend (e.g., Supabase, Vercel, Lovable).
- Nate highlights his open community tool suite: open-brain-stack (context layer), OpenSkills (reusable agent instructions), OpenEngine (agent handoffs), and ringer (multi-agent cost-optimized execution).