Last updated: July 2026
To write prompts that use fewer tokens, drop filler and politeness like "could you please," use imperative voice, stop restating context the model already has, reference files and functions by name instead of pasting them, and cut redundant adjectives. A token is roughly four characters or 0.75 words, so every word you trim saves real tokens.
Every word you type into an AI coding tool becomes tokens, and tokens are what you pay for. Most of us write prompts the way we write Slack messages — polite, padded, and repetitive. Here is how to tighten them without losing meaning, why it matters more in a long session than in a single message, and a before/after that shows the savings.
A token is the unit models read and bill by. As a rule of thumb, one token is about four characters, or roughly 0.75 words — so 100 words is around 133 tokens. Short common words are usually a single token; longer or rarer words split into several. This is why a wordy prompt costs measurably more than a tight one, and why the fixes below translate directly into fewer tokens sent and paid for. If you want to see the count for a specific prompt, our token calculator shows it before you hit send, and what token optimization is covers the broader idea.
The model does not need pleasantries to do the work. Phrases like "could you please," "I was wondering if you might be able to," "if it's not too much trouble," and "thanks so much in advance" are pure token overhead. The same goes for hedging — "I think maybe," "it would be great if," "perhaps you could." Strip them. "Could you please refactor this function to be more readable?" becomes "Refactor this function for readability." Same instruction, roughly half the tokens.
Commands are shorter than requests. "I would like you to add error handling to the login route" is a request wrapped around an instruction; "Add error handling to the login route" is the instruction. Imperative verbs — add, fix, rename, explain, refactor — start the sentence, carry the intent, and cut every word spent softening the ask. Models respond just as well to a direct command as to a polite one, and often more precisely.
If the model already has a file open, the conversation history, or a system prompt describing the project, do not re-describe it. "As you know, this is a React app using TypeScript and we're using Tailwind for styling, and in the previous message you saw the component…" is tokens spent telling the model what it can already see. Trust the context. Say what changed or what you want next, not a recap of everything leading up to it.
Pasting a 200-line file into your prompt to ask about one function can add thousands of tokens. If the tool can read your files — as Cursor, Claude Code, and most agents can — reference the file or function by name instead: "Fix the off-by-one in parseRange() in utils/date.ts." The agent reads exactly what it needs. When you must paste, paste only the relevant lines, and trim stack traces to the frames that matter. This single habit saves more tokens than all the wording tweaks combined.
A terse bullet list carries requirements with fewer connective words than flowing prose. Instead of "First I'd like you to add validation, and then after that it would be good to also write some tests, and finally please update the docs," write three bullets: - add validation, - write tests, - update docs. While you're at it, cut redundant adjectives — "a really clean, simple, elegant, well-structured solution" says nothing "clean solution" doesn't. Adjectives that don't change the output are just tokens.
Heavy markdown — decorative headers, bold everywhere, nested tables — adds formatting tokens the model doesn't need to understand plain instructions. Keep light structure (a simple bullet list) when it aids clarity; drop the ornamentation when a plain sentence works. Finally, batch related asks. Three quick follow-ups in one message share a single context payload instead of re-sending the accumulated history three times. Grouping related changes into one prompt is one of the biggest per-session savings there is — for the full picture on where those tokens go, see our breakdown of AI coding agent costs.
Here's a real-world prompt tightened with the rules above. The meaning is identical; the token count is not.
BEFORE (~48 tokens) "Hi! I was wondering if you could please help me out. Could you take a look at the getUser function in the auth file and maybe add some proper error handling to it? That would be really great, thanks so much!" AFTER (~18 tokens) "Add error handling to getUser() in auth.ts."
That's roughly a 60% reduction on this one message — and in an agent session that prompt rides along in the context on every following turn, so the saving is paid back turn after turn, not just once.
Be honest about diminishing returns. On any single request, the context — open files, chat history, and tool output — almost always dwarfs your typed prompt. Trimming ten words off a message that ships alongside a 6,000-token file dump is a rounding error on that turn. If token cost is your real concern, controlling context (referencing files, closing irrelevant ones, scoping tasks) moves the number far more than word choice.
But prompt discipline still compounds. In an agent or Composer session, your prompt is re-sent with the growing history on every turn, so a tighter prompt is paid back on each turn that carries it forward. Over a 30-turn session those savings add up, and the habit of writing tersely tends to produce more focused output as a bonus. Techniques like telegraphic compression and NLP-based analysis push this idea further, dropping function words a model can infer.
Terse fixes typos, removes filler and politeness, and shortens phrases in the prompts you type into AI coding tools — on-device, zero latency, no API calls. Three modes from a light touch to aggressive. It applies every rule on this page for you, before the prompt is sent.
How Terse WorksDrop filler and politeness like "could you please" and "I was wondering if," switch to imperative voice, stop restating context the model already has, reference files and functions by name instead of pasting them, use terse bullet lists over prose, and cut redundant adjectives. A token is roughly four characters or 0.75 words, so every word you remove saves real tokens.
As a rule of thumb, one token is about four characters or roughly 0.75 words. So 100 words is around 133 tokens. Short common words are often a single token; longer or rarer words split into several.
On any single message the context — files, chat history, and tool output — usually dwarfs your typed prompt. But prompt discipline compounds: in an agent session your prompt is re-sent with the growing history every turn, so a tighter prompt is paid back on every turn that carries it forward.
Heavy markdown — decorative headers, bold, nested lists, tables — adds tokens the model does not need to understand plain instructions. Keep light structure like a simple bullet list when it aids clarity, but drop ornamental formatting when a plain sentence works just as well.