dsh-code-reading-coach

Reading a paper and understanding its open-source implementation are two very different skills. Papers have a fixed skeleton; repositories do not. Most developers and researchers open a new repo after finishing a paper and immediately freeze, unsure whether to start at main.py, look at the test suite, or trace a single function call. dsh-code-reading-coach is a DeepSeek Harness Agent preset that removes that paralysis by guiding you through a structured five-stage method, asking only one question and giving only one next step at a time.

The five stages—Anchor, Terrain, Entry, Core Mapping, and Close the Loop—progress from distilling paper claims into code-level questions, to confirming you know the language and framework, to tracing a minimal execution path, to reading core modules and mapping each claim to its implementation, and finally to running a test, hand-tracing a data flow, tweaking a single line, and paraphrasing the result in your own words. Each stage yields a concrete artifact (claims list, terrain map, execution-flow diagram, mapping table, study notes), and everything persists to a workspace notes.md file so the session can be resumed across conversations.

The preset adapts its walkthrough to the repository type (training framework vs. kernel library vs. model repo), offers three depth levels with a built-in Feynman checkpoint, and explicitly flags any contradiction between paper and code. It is ideal for students and researchers digesting AI/ML papers, for engineers tasked with understanding an inherited codebase, and for anyone who wants a repeatable, low-pressure way to go from "I cloned the repo" to "I can explain the whole system."