Vision

Vision

Claude Code ships with a single system prompt designed to be safe and useful across all contexts. This is the right default — but it's a compromise. Instructions that make Claude careful and minimal during a surgical bug fix actively prevent it from building proper abstractions in a new project. Instructions that suppress verbose output hurt exploration sessions where you want Claude to explain its reasoning.

`claude-mode` solves this by replacing the behavioral layer of Claude Code's system prompt while preserving everything else — tool instructions, security guidelines, environment info, hooks handling — intact. It's a thin launcher, not a fork.

The Axis Model

Instead of a handful of hardcoded personas, `claude-mode` composes behavior from three independent axes:

**Agency** — How much initiative should Claude take?

• Autonomous: Creates files, restructures code, makes architectural decisions without asking. Suited for building and refactoring.
• Collaborative: Explains reasoning, checks in at decision points, confirms before large changes. Suited for unfamiliar codebases and working with less experienced developers.
• Surgical: Executes the specific request with minimal blast radius. Does not touch adjacent code.

**Quality** — What standard should the code meet?

• Architect: Proper abstractions, error handling at boundaries, forward-thinking structure. Code should be maintainable long-term.
• Pragmatic: Match the existing codebase's patterns. Improve incrementally where it makes sense, but don't impose a different paradigm. Favor consistency over perfection.
• Minimal: Smallest correct change. No refactoring, no new abstractions, no speculative improvements.

**Scope** — How far can Claude reach beyond the immediate request?

• Unrestricted: Free to create new files, modules, test suites, configuration. Can reorganize project structure.
• Adjacent: Can make related changes in the neighborhood of the request — fix a broken import, update a test, rename for consistency. Won't restructure unrelated code.
• Narrow: Only what was explicitly asked. Nothing else.

Presets

Seven named presets cover common workflows:

Preset	Agency	Quality	Scope	When to use
create	autonomous	architect	unrestricted	Building from scratch — proper structure, abstractions, and architecture
extend	autonomous	pragmatic	adjacent	Extending an agent-coded or fast-built project — improve quality incrementally, clean up as you go
safe	collaborative	minimal	narrow	Surgical changes to production code — minimal risk, maximum precision
refactor	autonomous	pragmatic	unrestricted	Restructuring — move files, consolidate modules, improve patterns across the codebase
explore	collaborative	architect	narrow	Understanding a codebase — read, explain, suggest, but don't change files
debug	collaborative	pragmatic	narrow	Inves