Harness Engineering Framework

Problem

When AI agents work on tasks that span multiple files, sessions, or unfamiliar systems, they can drift from context, misread existing behavior, introduce regressions, or over-specify brittle plans. In production codebases, jumping straight to implementation can turn small changes into hard-to-trace failures.

How it works

The skill enforces a grounded sequence: inspect the repo, plan the change, decompose the work, implement atomically, verify immediately, and review adversarially.

• Reconnaissance: read CLAUDE.md, AGENTS.md, .cursor/rules, or equivalent; check git status and recent commits; load progress.md, feature_list.json, or similar handoff artifacts.
• Planning: map entry points, data flow, module boundaries, invariants, and verification surfaces, then break the work into small, testable, reversible steps.
• Execution: re-read target code before editing, prefer additive changes, and limit impact to public interfaces, serialization formats, dependencies, and build configuration.
• Verification: run linters, tests, real API calls, browser automation, or screenshots; do not mark work complete based only on code review.
• Review: inspect the change from reviewer, QA, and architect perspectives, covering edge cases, silent failures, broken contracts, and consistency with existing patterns.

Boundaries

It is suited to large refactors, multi-session features, complex integrations, and high-risk changes. It should be skipped for single-file fixes, documentation-only edits, or simple configuration tweaks. When used, repository conventions take precedence, feature lists should be append-only, and acceptance criteria should not be softened to make checks easier to pass.