Harness Wise

Build the harness before coding. A harness is a compact operational brief that tells a coding agent what the task is, what context matters, which project patterns and skills apply, what risks exist, what not to do, and how to proceed safely.

This skill prepares, reviews, or explicitly saves that harness. It does not implement application changes.

Core Rule

Do not code while using this skill. If the user asks for implementation, first produce the harness or review the existing harness, then hand off to the appropriate planning or work flow.

Harness files are allowed when the user explicitly asks to create, write, save, generate, or update a harness artifact. A harness artifact is a markdown operational brief for another agent, not project application code. When "generate files" is mentioned in the context of this skill or a harness, assume the requested files are harness markdown files unless the user clearly asks for application code.

When the user did not ask for a file, proactively propose saving the harness if persistence would add value: long or cross-cutting work, multiple agent handoffs, work that may resume later, reusable repo intelligence, or a harness likely to be reviewed before execution. Propose a concrete path and explain why, but do not write the file until the user asks or confirms.

Do not claim or imply that this skill is included in the model's startup context, system prompt, default behavior, or always-on environment. Treat it as an explicitly invoked workflow only.

Do not add this skill itself to any generated or reviewed harness. The harness may recommend other task-relevant skills, but it must not instruct the next agent to use $krt:harness-wise, mention that this skill generated the harness, or embed this skill as part of the handoff.

Workflow