claude-better
Skill by ara.so — Daily 2026 Skills collection.
claude-better is a compatibility-first reimplementation of the Claude CLI focused on aggressive performance improvements: up to 73% faster startup and up to 80% lower resident memory, while maintaining 100% command-level compatibility with the original Claude CLI.
What It Does
- Faster cold starts:
--helpgoes from 182ms → 49ms;chatsession bootstrap from 311ms → 102ms - Lower memory: sustained interactive sessions drop from ~412MB → ~83MB RSS
- Drop-in compatible: 100% pass rate on primary command forms, 100% exit-code match, 98.7% byte-for-byte output parity
- Zero migration cost: existing scripts, aliases, and muscle memory continue to work unchanged
Availability
⚠️ Source code is provided for selected high-profile customers only and available upon request. Contact the maintainer at krzyzanowskim/claude-better for access.
If you have access, install as described in your onboarding materials. The binary is a drop-in replacement — substitute it wherever you invoke claude.
Installation (Once You Have Access)
Key Commands
claude-better mirrors the Claude CLI surface exactly. All commands you know work as-is:
Configuration
claude-better reads the same configuration as the original Claude CLI. No new config format is required.
Performance Characteristics
Scripting Patterns
Since compatibility is 100%, all existing scripting patterns work unchanged:
Compatibility Notes
- CLI surface: 100% compatible with targeted Claude CLI command forms
- Exit codes: 100% match on documented exit-code behavior
- Output parity: 98.7% byte-for-byte; 100% semantic parity after whitespace/timestamp/terminal-width normalization
- Tested environments: macOS (Apple Silicon), Linux, containerized CI
- Tested against: 1,200 synthetic invocations, 87 flag combinations, 42 interactive flows, 14 failure-mode scenarios
Troubleshooting
Binary not found after install
Unexpected output differences
Auth not recognized
Falling back to original CLI
Architecture Notes (For Contributors / Evaluators)
The performance gains come from specific implementation choices documented in the README:
- Zero-copy streaming pipeline for token output (reduces streaming jitter)
- Precomputed command registry instead of dynamic startup discovery (cuts cold-start time)
- Aggressively bounded allocation for session state (drives memory reduction)
- Lazy subsystem initialization — only the active command path pays startup cost
- Compatibility shim layer that preserves flags/behavior without carrying the full original stack


