Framework
The Four-Category AI Coding Taxonomy
AI coding tools fall into four distinct categories with different users, capabilities, and failure modes: (1) generalist code assistants (Copilot, Cursor) for developers; (2) prototyping tools (Lovable, Bolt, v0) for fast front-end visualization; (3) lifecycle agentic flows (Devin) for engineering teams across the full SDLC; (4) MVP builders (Lio) for full-stack requirements-to-scaffold generation. Knowing which category you're in tells you what to expect, where you'll hit walls, and where to go next.
How it works
- 1Category 1 (Generalist Assistants): Copilot, Cursor — IDE plugins, tech-stack agnostic, help developers write faster. Developer-operated, context-dependent, no requirements layer.
- 2Category 2 (Prototyping Tools): Lovable, Bolt, v0 — prompt to clickable UI in minutes. Accessible to non-technical users. Excellent for front-end visualization; not designed to produce production-ready, scalable backends.
- 3Category 3 (Lifecycle Agents): Devin, Codegen — span the entire SDLC from writing to committing to reviewing to debugging. Built for engineering teams, not founders. Assume the developer already has context.
- 4Category 4 (MVP Builders): Lio — requirements elicitation → full-stack scaffold (front + back end) → structured handoff. Designed to produce a minimal viable product ready to scale, not a prototype ready to be thrown away.
- 5The key failure mode: using a Category 2 tool and expecting Category 4 output. Prototypes don't become products without requirements; skipping the requirements layer causes the 70% wall.
Hear it on the podcast
Most AI Prototypes Die Before LaunchListen to the episodeMore from this episode
Requirements-First DevelopmentThe CTO's core job is not writing code — it's translating business objectives into technical requirements, then ensuring the code solves the right problem. Start there. Generating code before generating requirements produces something technically functional that solves the wrong problem. The 70% wall is the symptom; deferred requirements are the cause. Requirements elicitation — guided prompting to extract what you actually need — is what separates MVPs from prototypes.The 70/30 Software StackSeventy percent of every new software application is commodity: authentication, user management, CRUD operations, data pipelines, notifications. It was never innovative. Every dev shop had internal boilerplates for it. The competitive advantage in software has always come from the 30% that's unique. In 2026, this matters more: any decent CRM can be built in a day. Automate the 70%; focus all human effort and genuine IP development on the 30%.
Get the next framework first
One email a week with the frameworks, tools, and founder stories from the podcast. Free.
More frameworks for AI & Automation
Trust Before AutonomyAI agents are trust-limited, not capability-limited. They can already do more than most organizations are comfortable delegating. The unlock isn't a better model — it's accumulated evidence of reliability. Trust is built through transparency (sources, confidence levels, disclosure), demonstrated consistency over time, and gradual expansion of autonomous scope as each new boundary proves safe. Autonomy follows trust; it does not precede it.The AI Factory EffectAI compresses delivery timelines — but scope expands to fill the gap. Teams don't get time back; they ship more. MVPs complete faster and then continue straight into features, architecture, and workflow automation. Same headcount. Same contract length. Dramatically more output. The promise was leverage. The reality is a factory. Plan your workload, pricing, and client expectations around this dynamic, not the old assumptions.Metacognitive AI CollaborationThinking deliberately about how you're thinking — what to keep in your brain vs. what to outsource to AI, and what questions are worth asking when you can get an answer to any question in two seconds. The foundational AI collaboration skill. Prompting can be trained in an afternoon; metacognition is what separates people who get better through AI from people who get worse.Process Redesign vs. Process AutomationDon't automate your existing workflow. Ask what process you'd design from scratch if you could do things that weren't previously possible. The difference isn't speed — it's architecture. Existing processes are designed around human cognitive limits (six to seven post-it notes, six to seven ideas). AI-native processes blow past those constraints: 300 customer needs, 400 product ideas, synthetic testing across all of them, human judgment reserved for the handful that survive filtering.The Collaborative Results IndexMeasure human-AI collaboration on three axes — not just usage. Results: what's the actual impact? Relationship quality: are humans genuinely engaging as collaborators, asking good questions, thinking critically? Resilience: is the human developing domain expertise through the collaboration, or outsourcing judgment and atrophying? The third dimension is the dangerous one: the cognitive muscle you don't train, you lose.The Intelligent Product EngineA three-part architecture that makes AI generation predictable and enterprise-appropriate: context archetypes (different rule sets for different modes), real-time artifact monitoring (Figma and git sync to structured artifacts for humans and AI), and a rules engine (WCAG, brand governance, legal gates, release criteria). Each part is necessary; together they turn generative AI from a prototype toy into a production tool.