tsm v1·TS-Code·TS-Design·TS-Docs

Ship better code.Burn fewer tokens.

For AI workflows that need stronger first‑pass output.

TokenSmoker compiles messy prompts into cleaner, structured briefs — preserving intent while reducing rewrite loops and prompt noise.

Install TokenSmoker Documentation

14-Day Free Trial · No Credit Card · Upgrade Anytime

tokensmoker

Input

“Here’s my entire project, every prior prompt,

and all related files… [thousands of tokens,

constraints buried in noise]”

Compiling...

Compiled prompt

“Here is the exact structure, constraints, and

intent this task requires — nothing else.”

Preserved

✓engineering constraints

✓stack context

✓implementation intent

===== COMPILE SUMMARY =====

What improved

Structured brief — engineering intent and constraints intact

Likely outcome

Fewer correction prompts

Prompt delta

607→188tokens↓ 69%

Built for Modern AI Workflows

01Why TokenSmoker?

The real cost
isn’t the token bill.

The token bill usually isn’t the problem. The real cost is repeated corrections, lost constraints, and burning time fixing outputs your prompt already explained.

The problem

Models misread prompts

Dropped constraints, ignored exclusions, and missed deliverable shape all come from prompt noise — not model capability.

Rewrite loops compound quickly

A bad first pass costs more than the initial token bill. Re-prompting, clarifying, and correcting erases the time savings.

Generalized compression breaks things

Stripping tokens without understanding prompt structure destroys load-bearing content. The model gets shorter input and worse output.

Prompts aren't all the same

Code, design, and docs prompts have different structures. A single compiler treats them all wrong.

The solution

Instead of

“Here’s my entire project, every prior prompt, and all related files… [thousands of tokens, constraints buried in noise]”

TokenSmoker sends

“Here is the exact structure, constraints, and intent this task requires — nothing else.”

Constraints preserved

Fewer rewrites

Stronger first pass

02Example Compiles

What gets sent to the model.

Harness-specific compilers strip noise while preserving exactly what the model needs to produce a strong first-pass output.

TS-Design—Design compile

Before compile

1,254 tokens

// Raw design prompt — uncompiled

// 1,254 tokens

// Component context, prior iterations,

// inline style tangents, duplicate props,

// version history, redundant copy notes,

// vague intent without deliverable shape...

// Topology unclear. Exclusions buried.

// Tailwind classes fragmented.

// Asset URLs referenced inconsistently.

After compile

415 tokens · ↓ 67%

// TS-Design compile — component brief

// 415 tokens

// Component: PricingCard

// Owner: pricing-section · parent: PricingGrid

// Tailwind: card-elevated px-6 py-8 rounded-lg

// Icon: /icons/checkmark-lime.svg

// Copy: "Start free trial" (quoted, exact)

// Exclude: hover animation on mobile

// Exclude: dark-mode variant (deferred)

What improvedPreserved component structure and design intent

Likely outcomeCleaner first-pass UI output

TS-Docs—Docs compile

Before compile

775 tokens

// Raw docs prompt — uncompiled

// 775 tokens

// Audience mentioned twice, inconsistently.

// Tone notes scattered across paragraphs.

// Exclusions buried mid-prompt.

// Deliverable shape never stated explicitly.

// Structure implied, not declared.

// Author intent diluted by hedging.

After compile

543 tokens · ↓ 30% · preservation-first

// TS-Docs compile — preservation-first

// 543 tokens

// Deliverable: onboarding guide, 800–1000 words

// Audience: senior engineers, no basics

// Tone: direct, no filler sentences

// Structure: intro → install → first compile

// Exclude: pricing section (handled elsewhere)

// Exclude: CLI flags beyond core three

// Intent: show value before showing depth

What improvedPreserved structure, constraints, and intent

Likely outcomeStronger first-pass draft with fewer rewrite loops

03Benchmarks

Better prompts are the metric.
Token reduction follows.

The goal is stronger first-pass output. Smaller prompts are often the result — not the objective.

~65%

TS-Design · token reduction

Measured on real Purchased User Prompts. Component topology, Tailwind classes, and quoted copy preserved throughout.

< 2s

Average compile time

Context resolution runs locally before each request. No latency added to the model call itself.

TS-Docs

Preservation-first. Reduction varies.

Some docs prompts compile significantly smaller. Others compile to roughly the same size — or slightly larger — because preserving structure, exclusions, and author intent is the correct call. Both outcomes are intentional.

775→54330% reduction

686→760preservation-first

Based on early internal testing on real Purchased User Prompts. Results vary by harness, prompt type, and deliverable structure.

04What It Does

Preserve what matters. Strip the rest.

TokenSmoker compiles prompts into cleaner, structured briefs that preserve constraints and reduce rewrite loops.

Preserves load-bearing constraints

Exclusions, component topology, engineering constraints, deliverable shape, and narrative intent are treated as protected — not stripped as noise.

Reduces rewrite loops

When the model gets a precise brief instead of a bloated context dump, the first-pass output requires fewer corrections.

Keeps follow-up room open

Compiled prompts are structured so follow-up turns can narrow scope cleanly without re-sending full context.

Lowers token burn when safe

Redundant preamble, repeated history, and irrelevant context are stripped — but only when doing so doesn’t destroy signal.

Dispatches to harness-specific compilers

Code, design, and docs prompts are routed to compilers that understand their structure. Auto-detect handles code and design. Docs is user-selected.

Why harness-specific compilation

Coding, design, and docs prompts do not share the same structure. Generalized compression treats all three identically — which means it destroys exclusions, component topology, engineering constraints, deliverable shape, and narrative intent.

Harness-specific compilers know what each prompt type needs to preserve. A design compiler protects component ownership chains. A docs compiler protects tone, audience, and deliverable shape. A code compiler protects stack context and the actual engineering ask.

Live harnesses

TS-Codelive

Preserves stack traces, file paths, tech stack, and the actual engineering ask.

Stack tracesFile pathsTech stackEngineering ask

TS-Designlive

Preserves component ownership topology, Tailwind classes, quoted UI copy, asset URLs, icons, and animations.

Component topologyTailwind classesAsset URLsQuoted copy

TS-Docslive

Preservation-first. Structure, audience, tone, exclusions, deliverable shape, and intent stay intact. In v1 you always choose TS-Docs explicitly — auto-detect covers TS-Code and TS-Design only.

StructureAudience & toneExclusionsDeliverable shape

TS-CADplanned

Planned harness for design-to-fabrication prompts.

TS-Agentplanned

Planned harness for multi-step agent task prompts.

05Who is it for?

Built for developers
who take AI seriously.

If your AI tool produces good-enough output on the first pass, you don’t need this. If it doesn’t, the prompt is usually the problem.

$ smoke code "fix this Express login route, missing validation"TS-Code

Developers in AI coding tools daily

You use Cursor, Copilot, or Claude daily. Prompt quality and first-pass accuracy directly affect how fast you ship.

→ constraint dropped · first-pass miss · re-prompting again

Engineers tired of rewrite loops

You've noticed the model dropping your constraints, ignoring exclusions, or producing output that misses the deliverable shape.

$ smoke design --pasteTS-Design

Designers prompting AI for UI

You write design prompts with specific component topology, Tailwind classes, and quoted copy. TS-Design was built for you.

$ smoke docs --file readme-rewrite.txtTS-Docs

Technical writers using AI for docs

Your prompts have structure, audience, tone, and exclusions that matter. TS-Docs is preservation-first by design.

1254 → 415 tokens · ↓ 67% · TS-Design

Teams where prompt spend adds up

When everyone compiles before sending, tighter prompts compound — fewer rewrite loops and leaner context on every run.

$ smoke code --file context.txtTS-Code

Engineers who want control over context

You want to define what the model sees and what it doesn't — not leave it to chance or tool defaults.

06Get Started

Up and running in under a minute.

Install once, activate with your email, then compile with smoke, tsm, or tokensmoker.

install & activate

01Install

$ npm install -g tokensmoker

# recommended

02Or install from GitHub

$ npm install -g github:TokenSmoker/tokensmoker

# always latest

03Activate

$ smoke activate --email you@example.com

# keyless — your TokenSmoker account email

04Confirm

$ smoke status

quick start

$ smoke code "paste or type your prompt here"

$ smoke code "fix this Express login route, missing validation breaks it"

$ smoke code --file prompt.txt # from file

$ smoke code --paste # from clipboard

$ smoke design --paste # design harness

$ smoke design --file mockup.txt

$ smoke docs --paste # docs — user-selected

$ smoke docs --file readme-rewrite.txt

Runtime Behavior

tokensmoker, tsm, and smoke are equivalent binaries
Activation is global per user — set once, works in every project
Keyless activation: email and TLS — credentials stored locally in ~/.tokensmoker
Auto-detect routes between TS-Code and TS-Design today
TS-Docs is user-selected in v1 — pass smoke docs explicitly

Auto-detect

smoke without a harness auto-detects between TS-Code and TS-Design. For docs prompts, pass smoke docs explicitly.

14-day free trial

Full access during the trial. Upgrade with smoke upgrade anytime. Full CLI details in Documentation.

07Workflow

How it fits into your workflow.

TokenSmoker is a multi-harness prompt compiler: it sits between your prompt and the model. The harness does the work — you just write the prompt.

You write a prompt

Write your code, design, or docs prompt as you normally would. Pass it via inline text, file, or clipboard.

smoke, tsm, and tokensmoker are equivalent binaries.

TokenSmoker compiles it

The prompt is routed to the right harness — TS-Code, TS-Design, or TS-Docs — which strips noise while preserving load-bearing content.

Auto-detect handles code and design. Pass docs explicitly.

Model gets a precise brief

The compiled prompt preserves the constraints, structure, and intent the model needs. Stronger first pass. Fewer rewrite loops.

Token reduction often follows — but it’s the outcome, not the goal.

Your Prompt → TokenSmoker Harness → AI Model

Stronger first pass. Fewer rewrite loops.

08Trial

Start free.
Upgrade when ready.

14-day free trial with no restrictions. Install, activate with your email, and start compiling prompts right away.

Install TokenSmoker Documentation See harnesses →

What’s included

Full feature access — no restrictions
No credit card required to start
Runs locally — nothing transmitted
Upgrade with one command when ready

Upgrade

$ tokensmoker upgrade

09Notes

Built to be trusted.

Privacy-first architecture

Context resolution runs on your machine. No source code, file contents, or project data is ever sent to TokenSmoker servers.

Activation is global per user

Run smoke activate once with your account email. It applies to every project on your machine. No per-project setup, no config file to commit.

Keyless activation

Credentials are negotiated over TLS and stored locally. Nothing to copy into env files or rotate by hand.

Tool-agnostic by design

TokenSmoker works at the prompt layer — not inside any specific tool. Compatible with Cursor, Copilot, Claude, and all major AI environments.

TS-Docs is user-selected in v1

Auto-detect routes between TS-Code and TS-Design. For docs prompts, pass smoke docs explicitly. This will be extended in a future release.

Legacy commands still work

Prior compile commands are still supported. New work should use smoke <harness>. Migration is optional and non-breaking.

Less noise.
Better output.

Preserve what matters. Multi-harness compilation — TS-Code, TS-Design, or TS-Docs (explicit).

Start free trial — 14 days free →

Documentation

No credit card required·Full access during trial·Cancel anytime