Will Writing Code by Hand Really Go Extinct?
October 12, 2025
An open-source project called smol developer hit a thousand GitHub stars in a day. The demo shows an agent using Anthropic’s Claude with a 100,000-token context window to plan tasks, create files, and ship a Chrome extension from one prompt. Watch it once and it is easy to ask, “Is programming over?”
What Smol Developer Actually Does
Smol developer works because it uses the huge context window in Claude, sticks to coding instead of trying to be a general agent, and keeps you in the loop. The README calls it a “personal junior developer,” which is exactly how it feels. You set direction, it drafts code, and you stay on the hook for review. Used well, it is just another way to manage an eager junior.
Why This Feels Threatening
When tools like this improve, it is tempting to think one senior engineer with a small fleet of agents can replace a whole team. That fear is not new. Matt Welsh, a former Harvard CS professor and Google Principal Engineer, wrote “The End of Programming” for ACM Queue before smol developer even existed.
Welsh is not chasing clicks. He taught at Harvard for more than a decade—Mark Zuckerberg took his class—and then spent another decade leading big systems teams in industry. When he says hand-writing code will fade, he is talking about the work senior engineers already do. They spend their time clarifying requirements, designing systems, reviewing code, and mentoring. Typing every line is already rare at the top of the ladder.
The Skills That Gain Value
Welsh argues that as AI handles more of the typing, two skills move earlier in a career. The first is product sense: being able to explain what should exist and why it matters. The second is feedback: reading machine-generated code and spotting missing edge cases, awkward abstractions, or security problems. Staff engineers already live on those skills. The difference is that everyone else has to learn them sooner.
How To Practise Now
Before you open the editor, jot a short note about the problem, the constraints, and the edge cases. When an AI assistant hands you a diff, read it the way you would read a teammate’s patch. Does it follow the design? Do the tests cover the right cases? Is the style something you are willing to maintain? Keep learning how the pieces of the system fit together so you can steer even when the agent writes the boilerplate.
Humans will still ship software. The job shifts from “person who types every line” to “person who decides what should exist and ensures it works.” Lean into that shift now and the next batch of tools becomes leverage instead of a threat.