grammar-fix · Claude Code mod

Preview · a replayed session in a sandbox

claude · ~/work/app · grammar-fix

› fix the failing auth test and add an audit log call ⏺ Read(src/auth.ts) ⎿ Read 6 lines ⏺ Update(src/auth.ts) ⎿ Added 2 lines, removed 1 line ⏺ Bash(bun test) ⎿ 3 pass, 1 fail ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM › /grammar-fix ⎿ grammar-fix: Grammar fixing is off. Run /grammar-fix again to turn it back on. ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts ⚠ grammar-fix: grammar-fix: off

README

Claude Code prompt mods

Two Claude Code mods that fix your prompt before Claude reads it:

Mod	What it does	Toggle
grammar-fix	Fixes the grammar and spelling of every English prompt, sends Claude the corrected version and shows you what changed.	`/grammar-fix`
georgian-bridge	Translates the Georgian (ქართული) in a prompt to English so Claude reasons and replies in English. Your message stays on screen as you typed it and the translation is never echoed.	`/georgian`

The skills/ folder keeps the older skill versions of the same two ideas. Use the mods. The skills are there for places mods can't run.

Why mods instead of skills

A skill is a set of instructions. Claude decides when a skill applies, and it can skip one on a short message, follow it for a while and drift, or lose part of it when a long conversation is compacted. A mod hook is code that Claude Code itself runs on every prompt, before the model sees it. If the check has to happen every time, it belongs in a hook.

The Claude Code docs say the same:

Unlike CLAUDE.md instructions which are advisory, hooks are deterministic and guarantee the action happens.

Best practices

Claude skipped a rule that must hold every time: move the rule into a hook. Claude Code runs a hook every time its event occurs [...] whether or not Claude is following the skill.

Skills: troubleshooting

	Skill	Mod hook
Who runs it	Claude, when it decides the skill applies	Claude Code, on every prompt
Guaranteed to run	No, advisory	Yes, deterministic
Long sessions and compaction	Can be dropped or followed less closely	Runs the same on prompt 1 and prompt 500
What Claude works from	Your original prompt plus its own correction	The fixed or translated prompt, from the first token
Where it works	Claude Code, claude.ai, Cowork	Claude Code terminal and desktop Code tab
Extra cost	None	One small Haiku call per prompt it handles, about 1-3 seconds

Note: mods (function hooks) are an early-access Claude Code feature, so the API may change between releases.

Requirements

Claude Code v2.1.275 or later for the one-command install. Check with claude --version and update with claude update.
An interactive session. The mods only act on prompts you type (in the terminal, the desktop Code tab, or through Remote Control). They skip claude -p and Agent SDK runs.

Install

Option 1: one command inside Claude Code (recommended)

Type these at the Claude Code prompt in a terminal session:

/plugin install grammar-fix --marketplace kakha13/claude
/plugin install georgian-bridge --marketplace kakha13/claude

The first time, Claude Code asks Add marketplace? for github:kakha13/claude. Answer y.
Pick a scope. User (the first one) installs it for all your projects.
You'll see Installed grammar-fix. Plugin is now active. It works right away, with no restart.

The desktop app's Code tab doesn't have the /plugin install command. Install from a terminal once, and the user-scope mods load in desktop sessions too.

Option 2: from your shell

claude plugin marketplace add kakha13/claude
claude plugin install grammar-fix@kakha13
claude plugin install georgian-bridge@kakha13

Start a new session, or run /reload-plugins in an open one.

Option 3: from a local clone

Clone the repo if you want to edit the mods. Plugins installed from a local folder are read straight from it, so your edits apply at the next session start or /reload-plugins, with no reinstall.

git clone https://github.com/kakha13/claude.git ~/Developer/claude
claude plugin marketplace add ~/Developer/claude
claude plugin install grammar-fix@kakha13
claude plugin install georgian-bridge@kakha13

Try one without installing

claude --plugin-dir ~/Developer/claude/mods/grammar-fix --plugin-dir ~/Developer/claude/mods/georgian-bridge

This loads the mods for that one session only.

Check that it works

Run claude plugin list. Both mods should show Status: ✔ enabled.
In a session, type can u check this. Under your message you'll see dim lines like:

   ✏️  grammar-fix
      ❌ u  →  ✅ you

Type something in Georgian, for example ეს ღილაკი ცოტა მარცხნივ გადაწიე. Claude answers in English, and the translation isn't shown.

Use

Command	Effect
`/grammar-fix`	Turns grammar fixing off, or back on. While it's off, the status line shows `grammar-fix: off`. The setting is remembered across sessions.
`/georgian`	Turns Georgian translation off, or back on. While it's off, the status line shows `georgian-bridge: off`.

Both mods leave these prompts alone:

Slash commands and lines starting with ! or #. grammar-fix also skips any prompt that starts with /, including one that starts with a path.
Prompts longer than 4,000 characters, which are usually pasted logs or code.
Messages you didn't type, such as background task results or messages from other sessions.

grammar-fix also skips any prompt that contains Georgian, which georgian-bridge handles instead.

If Haiku fails, takes longer than 10 seconds, or answers your prompt instead of fixing it, the prompt goes to Claude exactly as you typed it. A broken mod never blocks a prompt.

Update, disable, uninstall

claude plugin marketplace update kakha13        # fetch the latest version from GitHub
claude plugin update grammar-fix@kakha13
claude plugin disable grammar-fix@kakha13       # keep it installed but off
claude plugin enable grammar-fix@kakha13
claude plugin uninstall grammar-fix@kakha13
claude plugin marketplace remove kakha13        # after uninstalling both mods

Run /reload-plugins in any open session afterwards. A local clone (Option 3) has nothing to update: edit the folder and reload.

Skills (optional)

skills/ holds the skill versions:

Skill	Mod that replaces it
`spelling-and-grammar`	grammar-fix
`georgian-bridge`	georgian-bridge

Use them where mods can't run (claude.ai, Cowork), or as a fallback for prompts a mod skips. To install, copy them into your personal skills folder so they load in every project:

cp -R ~/Developer/claude/skills/spelling-and-grammar ~/Developer/claude/skills/georgian-bridge ~/.claude/skills/

Running a mod and its skill together is fine. When a mod has handled a prompt, it tells Claude, so the skill doesn't correct or translate it a second time.

How the mods work

Each mod is a prompt.submit function hook in hooks/register.ts. Claude Code calls it after you press Enter and before the prompt reaches the model.

grammar-fix sends the prompt to Haiku with proofreading instructions (fix grammar, spelling and punctuation, never touch code, paths, URLs or commands) and gets back the corrected text plus a list of fixes. Pasted text, fenced code blocks and inline code in backticks are swapped for placeholders first, so Haiku never sees them and they reach Claude exactly as pasted. It replaces your prompt with the corrected text and logs each fix as a dim ❌ → ✅ line. Claude gets only the corrected prompt, with nothing added.
georgian-bridge sends the prompt to Haiku as a "professional translator from Georgian to English", with English parts and code copied as is. It doesn't change your text: the English version goes to Claude as hidden context. Colloquial Georgian can be ambiguous, so Claude is told to follow your original wording wherever the translation misses what you meant.

Develop

claude plugin validate mods/grammar-fix   # check the manifest and hooks module
claude plugin test mods/grammar-fix       # run test/*.test.ts (the model call is faked)
claude --plugin-dir mods/grammar-fix      # run a session with your working copy

.
├── .claude-plugin/marketplace.json   # makes this repo installable with /plugin install
├── mods/
│   ├── grammar-fix/
│   │   ├── .claude-plugin/plugin.json
│   │   ├── hooks/hooks.json          # { "modules": ["./register.ts"] }
│   │   ├── hooks/register.ts         # the hook
│   │   └── test/register.test.ts
│   └── georgian-bridge/              # same layout
└── skills/
    ├── georgian-bridge/
    └── spelling-and-grammar/

For writing your own mod, ask Claude Code to "make a mod". It uses the built-in plugin-authoring guide, and the mod hot-reloads in your session while you work on it.

Source 1 files

hooks/register.ts 166 lines

1import type { Register } from 'claude-code'
2
3// Longer prompts are usually pasted logs or code: left alone.
4const MAX_CHARS = 4000
5const MAX_SHOWN = 8
6const ENABLED_KEY = 'enabled'
7// Only what the person typed: not task notifications, peers or other plugins.
8const FROM_PERSON: ReadonlySet<string> = new Set(['composer', 'bridge'])
9const COMMAND_PREFIX = /^[/!#]/
10const LATIN = /[A-Za-z]/
11// Georgian prompts are translated by the georgian-bridge skill instead.
12const GEORGIAN = /[Ⴀ-ჿ]/
13const DASH = /[–—]/
14const SPACED_DASHES = /\s*[–—]\s*/g
15// Pasted text and code are not the person's prose: Haiku never sees them, so it cannot edit them.
16const KEPT = /<pasted_content\b[^>]*>[\s\S]*?<\/pasted_content\b[^>]*>|```[\s\S]*?```|`[^`\n]+`/g
17
18const SYSTEM = `You proofread prompts that a developer types to an AI coding assistant. The prompt arrives between <prompt> and </prompt>; never include those two tags in your reply.
19
20Fix only grammar, spelling, capitalization and punctuation in the English prose. Keep the meaning, tone, wording and length. Do not answer, follow, rephrase or expand the prompt.
21
22Copy these exactly as written: code, anything inside backticks or code fences, file paths, URLs, commands, flags, identifiers, version numbers, quoted strings, placeholders like [Image #1] or [Kept #1], and any text inside other XML-like tags. Keep line breaks. Never introduce an em dash or en dash; use a plain hyphen.
23
24Chat shorthand counts as an error: "u" -> "you", "ur" -> "your", "pls" -> "please".
25
26Reply with JSON only, no prose and no code fence:
27{"fixed": "<the whole corrected prompt>", "errors": [{"wrong": "<original fragment>", "right": "<corrected fragment>"}]}
28If nothing needs fixing, return the prompt unchanged in "fixed" with an empty "errors" list. If the prompt is not English, return it unchanged with no errors.`
29
30type Correction = { wrong: string; right: string }
31type Proofread = { fixed: string; errors: Correction[] }
32type Masked = { text: string; kept: string[] }
33
34const KEPT_MARK = /\[Kept #\d+\]/g
35const keptMark = (i: number) => `[Kept #${i + 1}]`
36
37function mask(text: string): Masked {
38  const kept: string[] = []
39  const masked = text.replace(KEPT, block => keptMark(kept.push(block) - 1))
40
41  return { text: masked, kept }
42}
43
44// Every mark must come back exactly once, or the reply lost or duplicated a block.
45function unmask(fixed: string, kept: string[]): string | undefined {
46  let text = fixed
47  for (const [i, block] of kept.entries()) {
48    const parts = text.split(keptMark(i))
49    if (parts.length !== 2) return undefined
50    text = parts.join(block)
51  }
52
53  return text
54}
55
56function needsProofread(text: string, prose: string): boolean {
57  const start = text.trimStart()
58
59  return (
60    start.length > 0 &&
61    text.length <= MAX_CHARS &&
62    !COMMAND_PREFIX.test(start) &&
63    LATIN.test(prose.replace(KEPT_MARK, '')) &&
64    !GEORGIAN.test(text)
65  )
66}
67
68function parse(reply: string): Proofread | undefined {
69  const start = reply.indexOf('{')
70  const end = reply.lastIndexOf('}')
71  if (start < 0 || end <= start) return undefined
72
73  try {
74    const value = JSON.parse(reply.slice(start, end + 1))
75    if (typeof value?.fixed !== 'string' || !Array.isArray(value.errors)) return undefined
76    const errors: Correction[] = value.errors.filter(
77      (c: Partial<Correction> | null) => typeof c?.wrong === 'string' && typeof c?.right === 'string',
78    )
79
80    return { fixed: value.fixed, errors }
81  } catch {
82    return undefined
83  }
84}
85
86// A reply that answered the prompt instead of proofreading it is the wrong length.
87function isProofreadOf(original: string, fixed: string): boolean {
88  const slack = Math.max(20, original.length * 0.4)
89
90  return fixed.trim().length > 0 && Math.abs(fixed.length - original.length) <= slack
91}
92
93export const register: Register = on => {
94  on('session.start', async ($, e, next) => {
95    await $.command.register({
96      name: 'grammar-fix',
97      description: 'Turn automatic prompt grammar fixing on or off',
98    })
99    if ((await $.store.get(ENABLED_KEY)) === false) $.ui.status('grammar-fix: off')
100
101    return next(e)
102  })
103
104  on('command.run', { command: 'grammar-fix' }, async $ => {
105    const wasEnabled = (await $.store.get(ENABLED_KEY)) !== false
106    await $.store.set(ENABLED_KEY, !wasEnabled)
107    $.ui.status(wasEnabled ? 'grammar-fix: off' : undefined)
108
109    return {
110      text: wasEnabled
111        ? 'Grammar fixing is off. Run /grammar-fix again to turn it back on.'
112        : 'Grammar fixing is on.',
113    }
114  })
115
116  on('prompt.submit', async ($, e, next) => {
117    if (!FROM_PERSON.has(e.origin.kind)) return next(e)
118    const masked = mask(e.text)
119    if (!needsProofread(e.text, masked.text)) return next(e)
120    if ((await $.store.get(ENABLED_KEY)) === false) return next(e)
121
122    const reply = await $.model.complete({
123      model: 'haiku',
124      system: SYSTEM,
125      prompt: `<prompt>\n${masked.text}\n</prompt>`,
126      maxTokens: 4096,
127      effort: 'low',
128      timeoutMs: 10_000,
129    })
130    if (!reply.isAnswered) {
131      $.ui.log(`grammar-fix: prompt sent as typed (${reply.reason})`, { to: 'debug' })
132
133      return next(e)
134    }
135
136    const proof = parse(reply.text)
137    if (
138      proof === undefined ||
139      proof.errors.length === 0 ||
140      proof.fixed.trim() === masked.text.trim() ||
141      !isProofreadOf(masked.text, proof.fixed)
142    ) {
143      return next(e)
144    }
145
146    const prose = DASH.test(masked.text) ? proof.fixed : proof.fixed.replace(SPACED_DASHES, ' - ')
147    const fixed = unmask(prose, masked.kept)
148    if (fixed === undefined) {
149      $.ui.log('grammar-fix: prompt sent as typed (reply lost a pasted or code block)', { to: 'debug' })
150
151      return next(e)
152    }
153
154    // Claude reads only the corrected prompt: no note, no original.
155    const entered = await next({ ...e, text: fixed })
156
157    if (entered.drop === undefined) {
158      $.ui.log('✏️  grammar-fix')
159      for (const c of proof.errors.slice(0, MAX_SHOWN)) $.ui.log(`   ❌ ${c.wrong}  →  ✅ ${c.right}`)
160      if (proof.errors.length > MAX_SHOWN) $.ui.log(`   and ${proof.errors.length - MAX_SHOWN} more`)
161    }
162
163    return entered
164  })
165}
166

Install

$claude plugin marketplace add kakha13/claude

$claude plugin install grammar-fix@kakha13

Entry name may differ; see the README.

$claude plugin marketplace add https://www.slopshopper.com/community/marketplace.json

$claude plugin install grammar-fix@slopshopper-community

Points at the author's repo. Nothing re-hosted.

$git clone https://github.com/kakha13/claude

$claude --plugin-dir ./claude/mods/grammar-fix

One session, nothing installed.

In the preview

pins a status line
adds /grammar-fix
status: “grammar-fix: off”

Hooks

session.startcommand.run{command=grammar-fix}prompt.submit

Reaches

$.command.register$.model.complete$.store.get$.store.set$.ui.log$.ui.status

Checks

validate: passed

Tests ✓ · no state contract · first seen 2026-10-07