Coding by Hand: Reclaiming the Craft in the Age of AI
Miguel Conner is spending three months at the Recurse Center coding without AI to reclaim the deep technical understanding that automated agents often obscure. By building LLMs from scratch and practicing manual Python development, he aims to strengthen his 'mental fitness' and foundational computer science knowledge. He believes that this period of manual labor will ultimately make him a more effective and high-leverage engineer when he returns to using AI.
Key Points
- AI coding agents can hinder learning by making assumptions for the developer, leading to a weaker grasp of the codebase.
- The mental effort required to write code by hand is a vital 'cognitive workout' that builds professional mastery and craft.
- Deep foundational knowledge of computer science and language models provides developers with more leverage when they eventually use AI tools.
- Collaborative, self-directed environments like the Recurse Center are ideal for building technical intuition through pair programming and diverse exposure.
- Mastering low-level skills, such as terminal-based debugging and manual file editing in Vim, increases overall developer efficiency and independence.
Sentiment
The community overwhelmingly agrees with the article's core thesis that manual coding builds irreplaceable skills and understanding. While there is genuine respect for AI as a supplementary tool, the dominant sentiment is skepticism toward heavy AI reliance and concern about deskilling. A minority voice advocates for embracing AI-driven workflows as the inevitable next evolution, but these views receive more pushback than support in the discussion.
In Agreement
- Teaching with constrained tools like line editors and assembly language forces students to internalize code and think before writing, building stronger mental models
- Writing code manually creates an essential mental map of the codebase that gets lost when AI agents write the structure, leading to confusion during maintenance
- Simpler tools are a forcing function for simpler, more maintainable code — without code search, auto-refactoring, or AI, developers write more legible and well-structured programs
- AI autocomplete was a better middle ground than agentic workflows because it kept the developer in a tight feedback loop with the code
- Concerns about the next generation of engineers lacking foundational skills are legitimate, as over-reliance on AI prevents building deep understanding needed for debugging complex systems
- The enjoyment and craft satisfaction of programming disappears when AI does most of the work, turning developers from craftspeople into foremen
- Software quality was already on a downward trend and AI-generated code risks accelerating that decline
Opposed
- Agentic coding has brought back the excitement of building software by creating a closer connection between thought and result, freeing developers from framework limitations
- AI tools are just another step in the natural progression of abstractions, similar to the shift from assembly to compiled languages
- For business contexts, demanding to understand every line of code reduces velocity compared to teams leveraging AI fully
- Spec-driven development with AI actually increases time spent thinking about desired outcomes rather than hacking around framework limitations
- Making things painful for students just because it was painful in the past is not necessarily productive learning — constraints can put people off rather than teach them
- Experienced developers who review AI output can produce production-quality code efficiently, with the development harness being the main investment