This AI Replaced My Keyboard

Engineer Cole Medin presents a provocative future: he claims over a year without writing a single line of code himself. An engineer his entire life, Medin now fully outsources his development process, relying exclusively on AI coding assistants. His system, detailed in the video "Claude Code Skill I use to drive my entire development process," functions as a radical vision for developer augmentation.

At its core, Medin’s method revolves around Skills. These are not intricate software packages; instead, each Skill represents a simple, reusable prompt. They function as a curated workflow, guiding an AI agent like Claude Code or Codex through a specific development task or process. Medin built a comprehensive library of these processes, which he now shares.

His philosophy emphasizes a minimalistic, plug-and-play approach. Unlike rigid frameworks such as GitHub Spec Kit or Gas Town, which demand full adoption, Medin’s Skills library augments existing developer workflows. Developers pick, choose, and customize individual Skills, integrating them seamlessly to enhance their current tools without a full paradigm shift.

Medin's system operates on a lean two-loop AI cycle, channeling his project management expertise into clear AI directives. This framework divides work into high-level strategy and granular execution, leveraging AI for both.

First, the Outer Loop focuses on strategic project definition. Medin, acting as a product manager, uses the plan-create-prd Claude Code Skill to outline the 'what' and 'why,' generating a comprehensive product requirements document. This skill then guides the AI in crafting the technical specification, detailing the 'how.'

Concluding the Outer Loop, the piv-slice-epic skill takes the PRD and technical spec. It meticulously breaks down the large project into discrete, dependency-aware tickets, ready for AI processing. This ensures a structured, manageable workflow for complex initiatives.

Next, the Inner Loop handles the day-to-day development grind. For each individual ticket, the AI takes charge: it plans the implementation, writes the code, and validates its output. The developer transitions into a critical supervisory role, reviewing and guiding the AI's focused execution.

Medin’s system, a collection of skills, fundamentally diverges from monolithic AI coding frameworks like GitHub Spec Kit and Gas Town. These impressive, yet highly opinionated, projects typically demand full adoption of their specific software development life cycle (SDLC), defining every step from initial concept to deployment. Medin’s approach, conversely, champions a minimalistic, plug-and-play philosophy.

This modularity presents a clear advantage, especially for experienced developers with established workflows. Rather than forcing a complete overhaul, Medin's library enables engineers to select individual skills—perhaps just the plan-create-prd skill for defining the 'what' and 'why'—and integrate them without disruption. Developers maintain agency, augmenting their personalized processes instead of being constrained by a rigid, all-encompassing system.

Customization stands as a core tenet of this flexibility. Each Claude Code Skill functions as a simple, reusable prompt, making it remarkably easy to edit and adapt. Teams can fine-tune these prompts to match their unique codebase, adhere to specific style guides, or align with existing documentation standards, such as storing technical specs in Confluence versus local Markdown files. For detailed technical specifications and further customization options, explore the Claude Code Docs.

Medin’s system culminates in a validation-first approach, which he identifies as "the most important thing in the entire process." This critical methodology liberates the developer from constant oversight, allowing the AI to "cook" autonomously while maintaining confidence in its generated outputs. It fosters a robust feedback mechanism, ensuring the developer remains the ultimate arbiter of quality and correctness for every line of AI-generated code.

Developer's role fundamentally transforms from a hands-on coder to a high-level architect and meticulous quality assurance lead. This means defining high-quality inputs — comprehensive PRDs and detailed technical specifications — that precisely articulate the project's 'what,' 'why,' and 'how.' The human effort shifts entirely to intelligent design and rigorous output validation, moving beyond the manual grind of implementation.

This re-orientation unlocks a profound productivity implication. A single developer can now orchestrate a significantly larger scope of work, effectively managing multiple projects or complex features simultaneously. By offloading the tedious, repetitive tasks of code generation and initial implementation to the AI agent, human creativity and problem-solving capacity are strategically redirected towards higher-order challenges, architectural strategy, and innovative design.

What are Cole Medin's AI 'skills' for Claude Code?

They are a collection of reusable, minimalist prompts that guide an AI coding assistant through specific development tasks, like creating a PRD or planning code implementation. They act as pre-defined workflows.

How does this approach differ from frameworks like GitHub Spec Kit?

Medin's system is modular and 'plug-and-play,' allowing developers to adopt individual skills into their existing workflow. Frameworks like GitHub Spec Kit or Gas Town are more monolithic, often requiring adoption of their entire development life cycle.

What is the 'two-loop' AI coding workflow?

It's a framework consisting of an 'Outer Loop' for high-level planning (PRDs, specs, epic slicing) and an 'Inner Loop' for the day-to-day coding on individual tickets (planning, implementing, and validating code).

Can these skills be used with AI assistants other than Claude Code?

Yes. While designed for Claude Code, Medin states the skills repository can be given to other agents like Codex or GitHub Copilot with instructions to adapt and install them for that specific agent.