Author rebuilds AI-assisted learning workflow after retention drop

On November 5, 2026, a writer published a post on cnblogs describing how he rebuilt his learning workflow around AI tools. He keeps 71 DDD column articles as Markdown files on his machine, 26 from one course and 45 from another, exported through geektime-dl. He describes the moment the download finished as feeling almost like having learned the material, but he is careful to say the two things share no causal link: without active processing and retrieval practice, the files are just 71 files. He can verify his doubt easily. He can open them at any time, yet he cannot say which layer should hold the boundary of an order aggregate.

The post cites a controlled experiment from Anthropic carried out in early 2026. In that study, 52 developers learned Trio, a Python async library they did not know. The group allowed to use AI tools scored an average of 50 on a follow-up test, while the group that wrote code by hand averaged 67. The AI group finished only about 2 minutes faster, a gap the study reports as not statistically significant. The author draws a wider conclusion from these findings: the cost of obtaining an answer has dropped sharply, while the cost of turning that answer into durable ability has not. He also notes that with answers so easy to obtain, people are less able to identify what they do not know.

Two sets of figures support his reading of the wider shift. Stack Overflow recorded 3,862 questions in December 2025, down 78% year on year, against a monthly peak above 200,000 back in 2014. In the same period, Stack Overflow's 2025 developer survey found 84% of developers were using or planning to use AI tools, up from 76% in 2024. The author is explicit about the limits of the comparison: question volume is not the same as learning activity, and the decline also reflects community moderation. His own case for the trend rests on how the act of posting a public question has changed. Posting on Stack Overflow, he notes, forces a person to spell out the environment, the minimal steps needed to reproduce the problem, what has already been tried, and the full error message. The new workflow he proposes has three parts, each tied to a ready-to-copy artifact: an input file, a prompt, and a weekly schedule. The source material ends there, without reporting how the revised workflow performed in practice.