🌊 vaib
Star on GitHub β˜…

Assembly for LLM Agents.
High-density DSL to save tokens and flow.

An open-source, zero-compilation linguistic wrapper and cognitive-steering protocol built for agentic terminals.

$

⚑ Why vaib?

Traditional LLM prompts waste tokens on conversational filler, imprecise descriptions, and redundant file rewrites. vaib is an open-source, zero-compilation linguistic wrapper and cognitive-steering protocol built for agentic terminals (claude-code, Cursor, VS Code Agents).

  • The 3 Zeroes: Zero Fluff (up to 85% token reduction via caveman mode), Zero Redo Tax (guard clauses eliminate hallucinated logic bugs), Zero Setup (runs natively in your agent context).
  • Prefix-Cache Optimized: Declarative frontmatter triggers 100% LLM cache hits.
  • Single-Line Friendly: Type rapid multi-step logic on a single line using inline ; chaining without messy multiline chat boxes.
  • Diff-Only Output: terse_diff saves 65%–75% output costs by emitting raw line mutations instead of full-file rewrites.

πŸš€ 60-Second Quickstart

Inject the core linguistic engine directly into your global agent profile:

vaib logic spec
curl -sSL [https://raw.githubusercontent.com/Chalarangelo/vaib/main/install.sh](https://raw.githubusercontent.com/Chalarangelo/vaib/main/install.sh) | bash

Fire up your agentic CLI terminal, type /skills (or verify ~/.claude/skills/vaib), and vaib becomes an active protocol layer.


πŸ“‰ Token Overclocking & Performance Metrics

Because vaib structures your inputs into high-density declarative data maps and skips conversational AI fluff, it leverages modern LLM prefix-caching to its absolute limits:

  • output_format: terse_diff – Slashes output costs by 65%–75% by forcing the agent to output raw structural Git line mutations instead of rewriting full files.
  • output_format: caveman – Achieves a brutal 85% text compression rate. The AI strips out verbs, pronouns, and polite greetingsβ€”speaking in pure, token-lean engineering fragments while maintaining 100% byte-exact code precision.

πŸ’Ž Semantic Toolkit

Stop writing paragraphs. Steer your agent with precise structural operators:

🚦 Flow & Logic Control

  • ; β€” Inline Chain. Express multi-statement workflows on a single line.
  • unless cond β€” Postfix Guard. Early-exit guard clause (throw: Err unless user.active?).
  • given: [a, b] β€” Preconditions. Logical assertion array before execution.
  • :symbol β€” Enum Keys. Concise status mapping (:pending, :approved).
  • method? β€” Inspection Suffix. Side-effect-free boolean checks.
  • action! β€” Enforcement Suffix. Hard database commits or runtime exception throws.
  • {{ loose thoughts }} β€” Natural Escape. Drop raw, unstructured thoughts mid-syntax.
  • "literal" / 'char' β€” Text Lock. Guarantees exact UI string or character preservation.

🎭 Cognitive Identity & Context Controls

  • &scope("path") β€” Context Guard. Restricts agent focus strictly to specified path(s).
  • &dry β€” Plan Lockdown. Forces a 3-line structural dry run before touching files.
  • &grill β€” Shifts agent into a hostile Principal Engineer hunting defects and security leaks.
  • &architect β€” Systems optimization mode. Forces schema trees and service boundaries first.
  • &mentor β€” Deep educational mode with pattern breakdowns and metaphors.

⚑ vaib in Action (Showcase & Examples)

See how vaib strips away the conversational noise and boilerplate, turning you into a high-leverage intent designer.

1. The Secure Payment Loop (Backend & Logic Architecture)

Instead of typing out a wall of text detailing database safety, retry limits, and errors, express your transaction layer in just a few structured lines.

vaib logic spec
---
vaib: Payments/Capture/v2
stack: [TypeScript, Prisma, Stripe]
output_format: terse_diff
---

# State: Wallet
- balance: decimal! [invariant: >= 0]
- status: enum(:active, :frozen, :flagged)

# Logic
- fn ProcessTransfer(from_id, to_id, amount) {
    from = Wallet.find(from_id) ; to = Wallet.find(to_id)

    # Postfix guards & enum symbols eliminate multi-line nested IFs
    throw: InsufficientFunds unless from.balance >= amount
    throw: FrozenAccount if from.status >== :frozen || to.status >== :frozen

    # Destructive action with backoff retry and fallback
    Stripe.charge!(from.customer_id, amount) [retry: 3]
    rescue: -> &ask("Gateway timed out. Fallback to local ledger?")

    from.balance -= amount ; to.balance += amount
    -> mutate: Wallet.save_all!([from, to])
}

2. The Dynamic UI Component (Text & State Preservation)

When building layouts, stop the AI from guessing your exact interface copy. Enforce strict character literals ('X') and string locks ("Text") to map Tailwind arrays cleanly.

vaib logic spec
---
vaib: Views/Dashboard/Alert
stack: [React, TailwindCSS]
output_format: terse_diff
---

# State: Banner
- variant: enum(info, warning, danger)
- active_tab: char [default: 'A']

# Logic
- fn RenderAlert() {
    match(variant):
      "danger" => 
        Text.label = "🚨 CRITICAL FAULT: Action Required Immediately!"
        Container.style = "bg-red-500 border-l-4 border-red-700"
"warning" => 
        Text.label = "Warning: Check system status metrics."
        Container.style = "bg-amber-400"

    Button.on_click -> 
        &trace -> {{ Dispatch telemetry, close banner, switch active_tab to 'B' }}
}

3. High-Leverage Non-Technical Steering (Ops & Automation)

vaib isn't just for writing raw code files. You can command your terminal agent to orchestrate deep diagnostics, handle telemetry issues, and translate the data for the rest of your company seamlessly.

vaib logic spec
---
vaib: Ops/IncidentTriage/v1
output_format: eli5
---

- Action:
    # Directly calls system logs and attaches tracking history
    tail("-n 100", "/var/log/nginx/error.log")
    &investigate("src/middleware/auth.ts")

    # Tell the agent to switch into a Principal Engineer mindset, audit its own finding,
    # and explain the fix like the reader is 5 years old.
    then:
      &grill -> &canary
      => &explain(to: "eli5")

🀝 Upstream Rule Synchronization

Persist custom shortcuts locally or share them upstream without leaving your agent terminal:

vaib logic spec
&memorize("Data/UltraCaching") -> &submit("Core/Syntax")

The engine distills execution history, builds a clean syntax snippet, stages a branch, and opens an upstream GitHub PR.


Stop coding. Start vibing. 🌊

Elevate your AI agent workflows with zero-compilation high-density notation today.

Get Started on GitHub β†’