nativish

Nativish

A writing coach plugin for Claude Code — corrects your English before every response. Built for non-native speakers who code with Claude.

Every message you send gets coached, inline, before the answer.

Install

/plugin marketplace add denys-petryniak/nativish
/plugin install nativish@nativish

Restart your Claude Code session for the SessionStart hook to take effect.

What it does

Before every response, Claude inspects your prompt:

• Real grammar, spelling, or word-choice mistakes → a corrected version plus a numbered list of fixes
• Clean prompt → a one-line compliment
• Slash command, short ack (yes, ok, thanks), or message in a non-Latin script → silent

Toggle:

• /nativish:off — disable for the conversation
• /nativish:on — switch to default mode (chat-forgiving)
• /nativish:strict — switch to strict mode (also flags missing apostrophes, lowercase first letter, missing terminal periods, and common abbreviations)
• Inline aliases: nativish:off / nativish:on / nativish:strict (or with a space: nativish off, etc.) — handy if you'd rather type than autocomplete. The marker must be the entire message, so pasting a doc that mentions one of them won't accidentally switch states.

Status markers tell you which state the coach is in at a glance:

• ✓ en-coach — default mode (active, chat-forgiving)
• ✓ en-coach (strict) — strict mode (active, flags chat-style typos too)
• ⏸ en-coach (off) — disabled

How it works

A SessionStart hook injects the English-coaching skill into the conversation's system context, so the rule applies on every user message — no manual skill invocation needed.

Privacy & security

Nativish is a self-contained Markdown + shell plugin:

• No network calls beyond the standard Anthropic API that Claude Code itself uses.
• No telemetry, no analytics, no external services.
• No system file modifications — the plugin only injects text into the conversation context via the SessionStart hook.
• No elevated permissions required — does not request bypass-permissions mode or any permission overrides.

The full installation is a few Markdown files and one shell script that cats them. Inspect everything in hooks/ and skills/english-coaching/.

Known limitations

• Model variance. The skill's output format (Modes 1/2/3) is enforced via prose rules, so compliance varies by model. Opus follows the rules reliably; Sonnet (the default for many setups) occasionally drifts — wrapping a Mode 3 marker in dividers, or skipping the coaching block on a long input. The adversarial fixture suite at tests/adversarial-prompts.md catches drift; expect a high-but-not-perfect pass rate on Sonnet.
• Pasted content gets coached too. Code, logs, error messages, quoted prose — all of it is treated as text to check. Expect occasional false flags on snippets you didn't write.
• Long prompts are truncated in Corrected:. Only the first 2–3 sentences are echoed back; fixes for later sentences still appear in the numbered list below.

Why "Nativish"?

A coined word — almost native. The plugin won't turn you into a native speaker, but it nudges your written English a little closer every conversation.

---

Made with Claude, for Claude. 🧡