5 Months of Real-World Harness

An AI engineer used Claude Code daily in production for 5 months, then built a workflow harness on top of it.

The core problem is that Claude Code is powerful but unstructured by default. It edits files, but lacks plan review before starting work, structured evaluation right before a PR, and the audit trail needed for team work, which led to continually doubting the results.

The newly built claude-code-harness is a workflow layer on top of Claude Code that places a human approval gate at every critical step. It doesn't move to the next step until the user explicitly presses "go."

It consists of the following:

• 16 skills (slash commands)
• 14 sub-agents with model routing
• 5 Node.js hooks
• path-scoped rules
• GitHub / Azure DevOps tracker adapters

Claude Code remains at the center of the design. Opus handles planning and judgment, Sonnet handles code writing, and Haiku handles data gathering, while an adversarial evaluator with a separate prompt actively looks for failure points in the execution results before a PR.

In the small-developer flow, /implement #42 reads a GitHub issue to create a plan, and once the user approves it, implementation and testing proceed in waves. After that, it goes through adversarial evaluation and drafts a PR, but the final merge always requires user confirmation.

In the enterprise flow, /story runs a 5-stage lifecycle, leaving behind artifacts such as brief.md, plan.md, test-strategy.md, evaluation.md, and acceptance, along with an acceptance contract.