A developer named Casey Primozic spent three months coding “by hand”—no IDE autocomplete, no snippets, minimal copy-paste, just a basic text editor and raw typing. He built games like a ray tracer and a Doom-style renderer from scratch. Hacker News lit up with 400+ comments debating if this masochistic ritual builds real skills or wastes time. Bottom line: it boosted his typing speed from 40 words per minute (WPM) to 85 WPM and sharpened his code memory, but skeptics call it theater in an AI-coding era.
The Experiment Details
Primozic documented his self-imposed rules on his blog “Handmade Hero,” a series where he live-codes a game engine. For three months in 2018, he disabled every convenience: no syntax highlighting beyond basic colors, no auto-complete, no multi-cursor editing, no stack overflow copy-paste beyond APIs. He typed entire functions from memory, even standard library calls. Goal? Force muscle memory for code patterns, cut cognitive load from tool-hunting, and hit “flow state” faster.
Results were measurable. Pre-experiment, his effective coding speed hovered at 40 WPM due to backspacing and lookups. Post-marathon, 85 WPM sustained over hours. He claims it let him hold 10,000+ lines of complex C code in working memory—ray marching shaders, matrix math, input handling. Benchmarks: his ray tracer rendered scenes at 60 FPS on 2015 hardware. No crashes from rote errors; he internalized syntax quirks like GLSL vector swizzling (.xyz).
Context matters. Primozic codes in C, a verbose language demanding precision. Average programmer typing speed sits at 60-70 WPM per keylogging studies (e.g., Monkeytype data), but coding drops it 30% with symbols and logic. Elite coders like John Carmack hit 120 WPM. Primozic’s jump aligns with deliberate practice research—10,000 hours style, but compressed.
Hacker News Breakdown
HN thread exploded: top comments split 60/40 pro-con. Fans praised it as anti-fragility training. One user: “Autocomplete hides bugs; typing exposes them.” Another shared a study—novices using IDEs write 20% buggier code initially (from ICSE 2016 paper on tool dependency). Detractors? “Pointless virtue signal. Real productivity comes from abstraction, not stenography.” Fair point: LeetCode speed contests reward pattern recall, but production code is 80% design, 20% typing per GitHub metrics.
Debate spilled into AI tools. With GitHub Copilot generating 40% of code (per their 2023 report), does “hand-coding” inoculate against over-reliance? One commenter cited a Microsoft study: Copilot users accept 30% suggestions blindly, introducing vulns. Primozic’s method forces ownership— you can’t blame the squiggle.
Implications for Coders Today
This matters because software dev tilts toward tool dependency. IDE market: VS Code holds 75% share (Stack Overflow 2023), laced with extensions that autocomplete 50% of keystrokes. Result? Junior devs ship faster but grok less—evident in rising rewrite costs, up 15% YoY per Stripe’s engineering reports.
Skeptical take: Primozic’s gains are real but narrow. Typing speed plateaus; his 85 WPM won’t fix architecture flaws. For finance/crypto/security, where audits catch 1-in-10K buffer overflows, muscle memory shines—manual crypto impls (e.g., hand-rolled AES) catch nonce reuse faster than generated code. But scale it? Teams don’t have three months; they need CI/CD pipelines.
Why try it? Start small: 30 minutes daily in vim with plugins off. Command:
:set nocompatible
:syntax off
:nnoremap p "0p # single paste only Track WPM with typometer. Expect frustration first week, fluency by month two. In crypto, where side-channel leaks kill wallets, owning your primitives beats AI hallucinations.
Net: Valuable for solos or kernel hackers, overkill for CRUD apps. HN’s hype underscores dev anxiety—tools empower, but at what skill cost? Primozic proves basics endure; ignore at your peril.
