Init Context Engineering

Goal

Turn a repository into a reliable working surface for future agents. Prefer a small set of durable, verified context files over long prompt dumps or stale documentation.

Workflow

1. Inspect the existing project before writing context.

   • Read root and nested AGENTS.md files, README files, package manifests, build configs, test configs, CI workflows, and obvious scripts.
   • Use rg --files first to map the repo.
   • Treat uncommitted user changes as owned by the user.

2. Identify the smallest useful context surface.

   • Use root AGENTS.md for repository-wide instructions.
   • Use nested AGENTS.md only when a subtree has different rules.
   • Use skills/<skill-name>/SKILL.md or .agents/skills/<skill-name>/SKILL.md only for reusable workflows that future agents should explicitly invoke.
   • Use extra docs only when the information is too large or too specialized for AGENTS.md.