⁹Here is a number that stopped me cold: 379 of 485 tokens in one working prompt were dead weight, roughly 78% of the text doing nothing at all. The prompt still returned clean answers every time. That is the trap, and it is costing you money on every call.
The number comes from a small, sharp tool that shipped on an engineering forum yesterday, built by a developer who calls it Token Sift. It lints your prompts for wasted tokens before you fire a single call. Free, MIT licensed, and it runs offline: no API key, no account, nothing leaving your machine.
That offline part matters more than it sounds: your prompts often carry customer data or internal rules you would never paste into a random web tool. Most of us never audit this, because the prompt works, so we assume it is fine. What follows is how the waste creeps in, what the linter catches, and a tight workflow to claw the money back.
In partnership with
The agentic era needs a different CRM. That’s Attio.
Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.
*Ad
How prompts get fat
Think about how a production prompt grows. Someone adds a rule after a bad output. Someone else pastes in a longer example, and a third person restates the tone guidelines just to be safe.
Nobody ever removes anything. So the prompt gets heavier, and every call carries that weight. I have watched this happen to my own prompts: they fill up like a junk drawer nobody cleans out.
Here is why I find this worth your time. The waste is invisible because the thing still works. A broken prompt screams at you.
A bloated one just smiles and sends you the bill.
What the linter catches
The tool runs 20 lint rules covering the usual suspects: repeated instructions, over-verbose JSON, random UUIDs pasted inline, and dynamic content sitting where it kills prompt caching. Each rule maps to a habit you probably have.
Repeated instructions are the classic. Always answer in JSON lives in the system prompt, again in the examples, and once more at the bottom. Say it once and move on.
Over-verbose JSON is sneakier. Pretty-printed objects with deep indentation, long key names, and null fields the model never reads still get tokenized. Whitespace is not free.
Random UUIDs are pure fat: a 36-character ID can eat a surprising number of tokens, and the model seldom needs the whole string to do its job. Trim it to the last chunk and nothing downstream notices.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
*Ad
The cache killer nobody notices
This is the one that got my attention. Put a timestamp or a user name near the top of a prompt and the cached prefix breaks on every call. You pay full price each time.
Nothing errors out, which is the whole reason it hides. The bill runs higher than it should, call after call, and no alert ever fires. Here is the fix: move anything that changes between calls to the very end, and keep the static instructions up front.
The linter also prints a cost estimate per 1,000 calls, and that number is the real feature. A wasted token on one call is nothing. Across a thousand calls a day, it becomes a line item you can take to a budget meeting.
Picture a support bot answering 5,000 tickets a day on a prompt that is 78% padding. You are paying for four wasted prompts to get one real one. That is not a rounding error; that is most of your inference bill.
How to run it without wrecking quality
The workflow is dead simple. Install it from npm, point it at your heaviest production prompt, not a toy example, and write down the per-1,000-call cost before you touch a thing. The prompt called most, or the one patched for a year, hides the most waste.
Then fix one category at a time. Start with cache killers and repeated instructions, since those pay the most for the least risk. Change wording, JSON, and caching all at once and you will never know which edit moved the cost or broke the answer.
Here is the move that makes it stick: wire it into CI. Because it runs offline, the linter can fail a build when someone ships a bloated prompt, the same way a code linter catches sloppy code. Pair it with a token budget per prompt file and the waste stops creeping back.
Spotting padding in a prompt is a habit before it is a tool, and it only sharpens with reps on real prompts. The claude and chatgpt tracks in 3 Minute AI build that eye in three-minute lessons, each ending with a task you run in a built-in chat lab. Daily lessons and two full courses are free.
Two cautions, both earned. Do not delete everything it flags on sight, because some repetition is deliberate emphasis that holds a model on track. Keep the old prompt in git, re-run, and spot-check real outputs before you trust the trim.
One honest note: the post has just a couple of upvotes, and that 78% figure came from a single prompt the author chose. Your numbers will differ. A tight prompt might show 10% waste; one patched for a year could show far more.
Unify Your Teams and Tech Stack With HubSpot
Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.
Why HubSpot and what's new
• Generate leads and automate marketing with Marketing Hub
• Build your pipeline and close more deals with Sales Hub
• Scale customer support and drive retention with Service Hub
• Keep customer data clean, connected, and actionable with one, unified platform
Join 306,000+ in over 135 countries using HubSpot to grow their businesses.
See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.
*Ad
Your heaviest prompt, before the next bill
Pull up the one prompt you fire thousands of times a day and run Token Sift on it before your next bill lands.
If you want the 20 rules, the cache-breaking traps, and the per-1,000-call cost math laid out step by step, read the full breakdown and cost estimate.
Worth 10 minutes if you want a real dollar figure to hand your budget-holder instead of a vague hunch that your prompts are bloated.
Poll: A linter says 40% of your busiest prompt is waste. What do you actually do?
Hit reply and tell us why.
