Vibe Coding Bundle

A collection of AI prompt engineering tools that lint, optimize, and compress prompts for better LLM interactions.

To know more about individual tools, see their own README.

Features

• 🔍 vibe-lint — Flags weak, vague, and contradictory prompt instructions with 17 lint rules
• 🎯 prompt-optimizer — Rewrites weak prompts into tight, effective ones using the Claude API
• 🗜️ token-squasher — Compresses verbose system prompts into dense, token-efficient instructions
• 📦 prompt-rules — The shared rule engine both linting tools import
• Three input modes everywhere — file arguments, piped stdin, or an interactive REPL
• CI-friendly — vibe-lint exits non-zero on error-severity findings and speaks --json

Prerequisites

• Node.js 18 or higher (ES modules, global fetch, AbortSignal)
• An Anthropic API key (for prompt-optimizer and token-squasher)

Installation

One command for all three tools

npm install -g vibe-coding-bundle
vibe

vibe opens a menu. Name a tool instead and its arguments pass straight through, so the launcher never stands between you and the work:

vibe lint my-prompt.txt          # lint
vibe opt my-prompt.txt           # rewrite with Claude
vibe squash system-prompt.txt    # compress

| Alias | Tool | What it does | | --- | --- | --- | | lint | vibe-lint | flags weak, vague and contradictory prompt instructions | | opt | prompt-optimizer | rewrites a weak prompt with Claude, and scores the result | | squash | token-squasher | compresses a verbose system prompt, keeps every rule |

vibe-coding-bundle is the same command, if you prefer the long name. --no-splash skips the opening animation.

Or install individual tools globally

npm install -g @sidhxntt/vibe-lint
npm install -g @sidhxntt/prompt-optimizer
npm install -g @sidhxntt/token-squasher

Or clone and run locally

git clone https://github.com/sidhxntt/Vibe-Coding-Bundle.git
cd Vibe-Coding-Bundle
npm install
npm run build      # compiles the launcher and copies the tools into dist/
npm test           # runs every tool's suite
npm run dev        # the launcher straight from src/, via ts-node

Configuration

export ANTHROPIC_API_KEY="your_api_key_here"

| Variable | Effect | |---|---| | ANTHROPIC_API_KEY | Required by prompt-optimizer and token-squasher. | | VIBE_NO_API_NOTICE | Silences the first-run "your prompt is sent to the API" notice. | | NO_COLOR | Disables colour (also disabled automatically when stdout is not a TTY). |

Your prompt leaves your machine

prompt-optimizer and token-squasher send the full prompt text verbatim to the Anthropic API. Nothing is redacted. Both print a one-time notice before their first request — do not paste API keys, credentials, or customer data.

Usage

Every tool accepts the same three input modes:

tool file.txt file2.txt      # file arguments, processed non-interactively
cat file.txt | tool          # piped stdin
tool                         # no args on a TTY → interactive REPL

vibe-lint

vibe-lint my-prompt.txt
vibe-lint --json prompts/*.txt | jq '.[].score'
vibe-lint --max-warnings 0 system-prompt.txt      # strict CI gate
echo "You are a helpful AI assistant." | vibe-lint

Real output from vibe-lint --no-suggestions --quiet src/lint/bad-prompt.txt (exit code 1):

  vibe-lint v2.0.1  ─  src/lint/bad-prompt.txt
  Flagging weak, vague, and counterproductive prompt instructions

  ✖ error [E005] Reference to context not present in the prompt — model will hallucinate
    1:75  "the code"
  ✖ error [E002] Conditional hedges let the model decide scope — it will decide wrong
    1:84  "as needed"
  ✖ error [E001] Ambiguous delegation — model will hallucinate scope boundaries
    1:172  "do your best"

  ──────────────────────────────────────────────────────────
  Found: 3 errors  │  7 warnings
  Score: 0/100  [░░░░░░░░░░░░░░░░░░░░]

Exit codes: 0 clean · 1 errors present or --max-warnings exceeded · 2 usage error or unreadable file.

prompt-optimizer

prompt-optimizer my-prompt.txt
prompt-optimizer --force --model claude-opus-5 my-prompt.txt

1. Analyzes the prompt with the shared rules
2. Shows the findings with severity levels
3. Sends prompt + findings to Claude (temperature 0.2, bounded retry, hard timeout)
4. Re-lints the rewrite and reports which rule IDs were resolved, which are still present, and which the rewrite introduced

--force optimizes even when no rule fires, so a genuinely weak prompt that trips no regex still gets help.

token-squasher

token-squasher system-prompt.txt
token-squasher --verbose --model claude-opus-5 system-prompt.txt

Token statistics are on by default and come from the Anthropic count_tokens endpoint, with the compression request's own billed usage reported separately from resp.usage.

Compression is not verified — nothing checks that the model actually preserved every rule. Diff the two prompts before shipping.

Commands: :stats · :verbose · :clear · :quit

Rule categories

| Category | Description | Examples | |----------|-------------|----------| | vague-quality | Unmeasurable descriptors | "make it better", "high-quality" | | ambiguous-scope | Unclear task boundaries | "handle it", "as needed" | | output-format | Unspecified response format | "good format", "respond appropriately" | | role-confusion | Redundant AI identity statements | "you are an AI assistant" | | no-examples | Trailing example placeholders | "such as:" at end of line | | contradiction | Conflicting instructions | "be brief … be comprehensive" | | politeness-bloat | Softening constructions | "please try to", "feel free to" | | negation-only | Prohibitions with no paired DO | "don't be vague" | | missing-context | References to absent context | "the code", "the file" | | no-success-criteria | Subjective stop conditions | "until it's good" | | chain-of-thought | Reasoning suppressed | "just give me the answer" | | vague-persona | Personas with no domain | "act as an expert" |

vibe-lint --rules prints the live list.

Project structure

Vibe-Coding-Bundle/
├── package.json               # the launcher package: one bin, merged deps
├── tsconfig.json              # compiles src/*.ts only — the tools ship as-is
├── src/
│   ├── index.ts               # launcher: menu, dispatch, argv passthrough
│   ├── tools.ts               # alias → entry point registry
│   ├── splash.ts              # the opening animation
│   ├── rules/                 # @sidhxntt/prompt-rules — the shared rule engine
│   │   ├── index.js           # RULES, analyzePrompt, scorePrompt
│   │   └── test/rules.test.js
│   ├── lint/                  # @sidhxntt/vibe-lint
│   │   ├── index.js           # CLI, renderers, JSON reporter
│   │   ├── good-prompt.txt    # fixture read by the tests
│   │   ├── bad-prompt.txt     # fixture read by the tests
│   │   └── test/cli.test.js
│   ├── opt/                   # @sidhxntt/prompt-optimizer
│   │   ├── index.js           # CLI + Claude API integration
│   │   └── test/optimizer.test.js
│   └── squash/                # @sidhxntt/token-squasher
│       ├── index.js           # CLI + REPL + token statistics
│       └── test/squasher.test.js
└── docs/                      # the GitHub Pages site

The three CLIs are unchanged standalone programs. The launcher starts each one as a child process with inherited stdio, so a tool owns the terminal — and its own REPL — exactly as it does when installed on its own. Each keeps its own package.json, which is where it reads its version from, and type: module there is what keeps the tools ESM inside a CommonJS launcher.

Adding custom rules

Rules live in src/rules/index.js and are imported by both linting tools — there is no second copy to keep in sync.

{
  id: "W012",                          // unique identifier
  severity: "warn",                    // "error" | "warn" | "info"
  category: "vague-quality",           // rule category
  pattern: /\b(make it pop)\b/gi,      // global regex
  message: "Brief explanation",        // problem description
  suggestions: [                       // specific alternatives
    "Specify the axis: 'reduce complexity'",
    "Target a metric: 'cut latency by 30%'",
  ],
  docs: null,                          // optional documentation link
  appliesTo: (text) => true,           // optional whole-document guard
}

Add a matching entry to CASES in src/rules/test/rules.test.js; the suite fails if a rule ships without a known-positive and a known-negative.

Severity levels

• Error — high-confidence problems (ambiguous scope, missing context, contradictions). They make vibe-lint exit non-zero.
• Warn — style problems and improvement suggestions. They affect the score and --max-warnings, not the default exit code.
• Info — general recommendations.

Severity does not gate prompt-optimizer: all three are fed to Claude identically. Errors only weigh more heavily in the score.

Examples

Before (bad prompt)

You are a helpful AI assistant. Please try to be more helpful and improve
the code as needed. Make it better and more professional. Feel free to handle
the above code and do your best.

After (the shape an optimized prompt should take)

You are a senior TypeScript engineer. Review this code for:
1. ESLint strict compliance
2. Functions >20 lines need JSDoc comments
3. Cyclomatic complexity <10
4. No magic numbers

Return: markdown with ## Issues and ## Fixed Code sections.

Contributing

1. Fork the repository
2. Create a feature branch: git checkout -b feature-name
3. Add rules to src/rules/index.js, following the existing pattern
4. Add a positive and a negative test case in src/rules/test/rules.test.js
5. Run the suites: npm test from the repo root
6. Verify the fixtures still behave: node src/lint/index.js src/lint/good-prompt.txt (exit 0) and node src/lint/index.js src/lint/bad-prompt.txt (exit 1)
7. Submit a pull request

Testing your changes

npm test                                   # every tool's suite
node --test src/lint/test/*.test.js        # one tool

node src/lint/index.js src/lint/bad-prompt.txt
node src/opt/index.js --help
node src/squash/index.js --help

npm run build && node dist/index.js lint src/lint/bad-prompt.txt

Publishing

vibe-coding-bundle vendors the rule engine, so it has no release order to observe: npm publish builds first via prepublishOnly and ships dist/ alone.

The individually published CLIs still depend on @sidhxntt/prompt-rules, so that package must go to the registry first; a release of @sidhxntt/vibe-lint or @sidhxntt/prompt-optimizer will not install for anyone until it is there.

License

MIT License — see LICENSE.txt at the root and in each tool directory under src/.