dsh-code-reading-coach

Reading an AI or systems paper is one thing; making sense of its companion open-source repository is another. Papers follow a predictable structure, but the code behind them varies wildly in language, framework, and layout, leaving most readers staring at a tree of files with no clear entry point. dsh-code-reading-coach is a DeepSeek Harness Agent preset designed for exactly this gap: it walks you through a five-stage method that turns "I have no idea where to start" into "I can narrate the whole system," one small question at a time.

The five stages are: Anchor, where the paper is distilled into three to five core claims that become the questions your code must answer; Terrain, where your familiarity with the language and framework is confirmed before scanning dependencies, directory structure, and test layout; Entry, where a minimal runnable path is identified and the top-level execution flow is traced; Core Mapping, where key modules are read in full and each paper claim is matched to its implementation; and Close the Loop, where you run a minimal test, hand-trace one data path, tweak a single line to observe behavior, and explain it back in a Feynman-style summary. Every stage produces a tangible artifact, and all of them accumulate in a workspace notes.md file so you can pick up where you left off in a later session.

The coach adapts its path to the repository type—training frameworks, kernel libraries, and model repos each get slightly different emphasis—and offers three depth levels (overview, trace, deep-dive) with a Feynman checkpoint before deep-dive to make sure you can actually explain what you just read. When the code contradicts the paper, the discrepancy is flagged explicitly, which is often the most valuable insight. It is well suited to researchers and students working through AI/ML papers, as well as engineers who inherit a bare repository and want a structured, repeatable way to understand it without a formal companion paper.