Why this matters
LLM-driven coding assistants often make silent assumptions, overcomplicate solutions, and edit code beyond the requested scope. That behavior increases review friction, introduces regressions, and wastes developer time. These guidelines convert those vague instructions into explicit heuristics and verifiable goals so coding agents behave more like cautious, test-driven engineers.
What Sets It Apart
- Explicit operator-style rules, not abstract principles — four concise rules map directly to concrete behaviors: state assumptions, prefer minimal implementations, touch only requested code, and require verifiable success criteria.
- Agent-first design — written to be installed as a Claude Code plugin or appended per-project (CLAUDE.md), so the guidance travels with the project and can be enforced by automation rather than relying solely on human review.
- Test- and goal-driven focus — emphasizes writing tests and defining success criteria before edits, enabling autonomous loops that are measurable and safe to run without constant human direction.
- Surgical edit discipline — rather than broad refactors, the guidelines force minimal changes and make the agent surface unrelated issues as notes instead of automatically fixing them.
Who It's For and Tradeoffs
Great fit if you use Claude/Claude Code or agent-based coding workflows and want fewer noisy diffs, clearer PRs, and safer autonomous edits. Teams that integrate LLMs into CI/CD or use agents (e.g., Cursor/Multica) will get immediate value by reducing brittle or overreaching changes.
Look elsewhere if your workflow expects agents to refactor and modernize large codebases proactively — these guidelines bias caution over aggressive improvement. Also, trivial one-line fixes may not need the full rigor; the guidance suggests pragmatic judgement for very small changes.
Where It Fits
This is a lightweight, behavioral layer rather than a feature library. It complements agent runtimes (Claude Code, Multica, Cursor) and prompt-engineering work: use the CLAUDE.md to codify project-specific constraints and let the agent follow measurable, minimal-change procedures.
By turning vague requests into concrete success criteria and surgical steps, the project reduces the class of mistakes that cause regressions while preserving agent productivity.
