Se-harness: AI coding tool unifies instructions across Claude, Gemini, Copilot

Se-harness addresses a fragmentation problem in AI-assisted development: as teams adopt multiple coding assistants, they accumulate duplicate instruction files. Each tool—Claude Code, GitHub Copilot, Gemini CLI, and others—reads from different paths (CLAUDE.md, .github/copilot-instructions.md, GEMINI.md) and maintains no shared awareness of one another, forcing developers to manually sync rules across seven or more files whenever a guideline changes. Se-harness solves this by introducing a layered resolution order: packages (optional, shareable git repos of versioned rules) take precedence over global rules (~/.seh/ on the user's machine), which override bundled core content. A project index (AGENTS.md + .seh/ modules) extends global rules but never contradicts them; running seh sync regenerates all tool-specific files atomically from these sources. This model also generalizes to Skills—reusable capability modules that route to each agent (always invoke, invoke-only-on-condition, or optional), eliminating duplication of prompts like debugging procedures or code-formatting checklists across teams.

The tool's design emphasizes avoiding vendor lock-in and keeping git as the sole external versioning system. Harness packages are plain directories, not proprietary containers; symlinks are used purely as pointers, and generated files are gitignored (not committed), so drift detection is straightforward. The new machine workflow—git clone a harness, seh package use it, seh package install --all—mirrors familiar git workflows, reducing onboarding friction for technical teams. By allowing teams to commit .seh/project.md and stack modules while ignoring generated CLAUDE.md and AGENTS.md files, the system keeps a single source of truth in version control without duplicating intent.