Macha

An opt-in response style for AI coding assistants: compact, natural, Tamil-influenced South Indian English. Macha shortens the assistant's replies—not your prompts, files, code, commands, or tool output.

Install

Every Agent Skills-compatible assistant:

npx skills add kingroryg/macha --skill macha --agent '*' --global --yes

Native installers:

# Claude Code
claude plugin marketplace add kingroryg/macha
claude plugin install macha@macha

# Gemini CLI
gemini extensions install https://github.com/kingroryg/macha

Use

Send /macha on to turn it on. Send /macha off to turn it off.

How it sounds

These are replies the model may compose—not transformations of the user's words.

Film references are garnish: at most one per reply, only when casual and low-stakes. They never replace the result, warning, uncertainty, or next action.

Paavam. means a sympathetic “poor thing” or “that's unfortunate,” never mockery or dismissal.

Macha keeps negation, conditions, numbers, units, code, commands, paths, URLs, identifiers, citations, quotes, and exact errors intact. Security warnings, destructive actions, serious topics, and formal artifacts stay in clear standard prose.

Stats

Thirty paired coding-assistant replies expressing the same intent used about 42% fewer output tokens in Macha:

This is an output-only, authored-pair benchmark—not a live-model A/B evaluation. The corpus excludes input, tool output, code, and skill-loading cost. The skill itself is 983–996 tokens and hosts may resend it each turn, so short or already-terse sessions can be net-negative.

python3 -m pip install -r requirements-benchmark.txt
npm run benchmark
npm run audit:phrases

For the live normal-vs-terse-vs-Macha harness and evidence rules, see benchmarks.

Development

npm test
python3 /path/to/skill-creator/scripts/quick_validate.py skills/macha

skills/macha/SKILL.md is the single source of behavior.