SKILL.md
$27
Installation (Once You Have Access)
# Typical binary drop-in replacement pattern
# Place the claude-better binary in your PATH before the original claude
export PATH="/path/to/claude-better/bin:$PATH"
# Verify it's being picked up
which claude
claude --version
# Or alias it explicitly without touching PATH
alias claude='/path/to/claude-better/bin/claude-better'
Key Commands
claude-better mirrors the Claude CLI surface exactly. All commands you know work as-is:
# Show help (cold start: ~49ms vs 182ms baseline)
claude --help
# Check auth status (warm start: ~58ms vs 146ms baseline)
claude auth status
# Start an interactive chat session (~102ms bootstrap vs 311ms baseline)
claude chat
# One-shot non-interactive command (~131ms vs 428ms baseline)
claude -p "Summarize this file" < input.txt
# All standard flags pass through unchanged
claude --model claude-opus-4-5 chat
claude --output-format json -p "List 3 facts about Rust"
Configuration
claude-better reads the same configuration as the original Claude CLI. No new config format is required.
# Standard Claude CLI env vars are respected
export ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY
# The tool reads ~/.claude/ config directory as normal
# No migration of config files needed
Performance Characteristics
Scenario
Baseline
claude-better
Improvement
--help cold start
182ms
49ms
73% faster
auth status warm
146ms
58ms
60% faster
chat bootstrap
311ms
102ms
67% faster
One-shot command
428ms
131ms
69% faster
RSS after 30min session
412MB
83MB
80% less
Streaming jitter p95
91ms
24ms
74% lower
Scripting Patterns
Since compatibility is 100%, all existing scripting patterns work unchanged:
#!/usr/bin/env bash
# Existing Claude CLI scripts work without modification
# Non-interactive pipeline usage
echo "Explain this error:" | cat - error.log | claude -p /dev/stdin
# Exit code handling (100% compatible)
if claude auth status; then
echo "Authenticated"
else
echo "Not authenticated — run: claude auth login"
exit 1
fi
# JSON output parsing
claude --output-format json -p "What is 2+2?" | jq '.content'
#!/usr/bin/env bash
# Long-lived interactive session — memory pressure is significantly reduced
# Useful on memory-constrained machines (laptops, CI runners)
claude chat
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
# Ensure claude-better/bin is earlier in PATH than original claude
echo $PATH | tr ':' '\n' | grep -n claude
which claude # should point to claude-better
Unexpected output differences
# 1.3% of outputs differ before normalization (timestamps, whitespace, terminal width)
# If a script breaks on exact output matching, add normalization:
claude -p "..." | tr -s ' ' | sed 's/[[:space:]]*$//'
Auth not recognized
# claude-better reads the same auth store as the original CLI
# If auth fails, re-authenticate via the standard flow:
claude auth login
Falling back to original CLI
# If you hit an edge case, unset the alias/PATH change to revert instantly
unalias claude
# or
export PATH="<original-path-without-claude-better>"
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