Harness engineering is a structured approach to AI-assisted development that replaces vague prompts with two disciplined phases: a repository impact map (where the AI inspects real code via LSP/MCP to produce a grounded plan reviewed by a human) and a strict Jira task template (with real file paths, symbol names, acceptance criteria, and test requirements). The core principle is 'structure in, structure out' — constraining the AI's solution space at each phase makes its output more predictable and targeted. Key practices include keeping conventions in the repo itself, versioning prompts and MCP configs as code, and tracing errors back to missing harness constraints rather than just fixing the output.
Table of contents
The problem: Vague in, vague outThe fix: A two-phase workflowWhy this worksThe takeawaySort: