skill-framework

When you let an AI assistant write code, the real failures are rarely syntax errors—they are process-discipline failures: skipping design and jumping straight to code, shipping without verification, debugging by guesswork, declaring "done" with no evidence. Praxis (dsh-praxis) distils the daily SOPs of a senior engineer into 14 triggerable skill cards so the model automatically "flips to the right page" when a matching scenario appears, turning gut feeling into structured workflow. The skill set spans requirement clarification, design conversation, implementation blueprint, step-by-step execution with checkpoint validation, test-first (red-green-refactor), evidence-based fault isolation, observable completion proof, parallel task splitting, delegated sub-agent builds with a two-round review gate, pre-review self-check and point-by-point feedback assimilation, same-repo branch isolation, branch conclusion and merge, all the way to skill authoring itself—forming a complete engineering loop. Every skill follows the Agent Skills open standard (SKILL.md + YAML frontmatter) and is registered through dsh's Cordis plugin interface as a skill provider; unloading cleanly removes all registration state with no residue. It is built for individuals or small teams already developing inside DeepSeek Harness who want their assistant to work by the book instead of improvising. After installation the model loads the matching skill automatically when a trigger scenario is hit; alternatively you can drop the skills/ directory straight into a dsh local skill root with zero code—same content, same effect.