Fill eight fields, get a prompt that survives the first retry, and see exactly what it costs at your volume. Twelve original templates, real BPE token counts, live model prices.
100% free · no signup · browser-local
Share on XTwenty task prompts, written for work people actually bring to a model. Pick one and replace the bracketed slots.
Executive brief: Turn a long document into a one-page decision brief a busy executive will actually read.
Every field is optional. Empty fields simply disappear from the prompt.
Chat reads naturally in a message. System keeps labeled sections for an API call.
Role: You are a chief of staff who writes for an executive with four minutes and no appetite for background. Task: Turn the source material into a one-page decision brief with a clear recommendation. Context: Source material: [PASTE THE DOCUMENT OR THE RELEVANT SECTION] Audience: A CEO who already knows the background and needs the decision, the evidence, and the risk. Constraints: - Lead with the recommendation, not the history. - Separate facts from interpretation. - Do not invent numbers, dates, or names. - Flag anything the source does not answer. Output format: A 40-word bottom line, then three bullets of evidence, then one open question. Length: Under 200 words.
0 tokens is <0.01% of Claude Sonnet 5.5's 1M context window. Cost uses list input pricing from the live catalog; output tokens are extra.
Checks the brief, not the wording. Placeholders are expected here.
If the format is unusual, add one short example of input → output.
Unresolved: [PASTE THE DOCUMENT OR THE RELEVANT SECTION]
Most prompt advice is a pile of tricks. Most prompt failures are simpler: the model was never told what the answer was for. This builder keeps the brief in separate fields so a missing part is visible before you send it.
Who the model should be for this task.
A role shifts vocabulary and judgment. "Staff engineer reviewing a diff" and "friendly technical writer" produce different reviews of the same code.
One line. Skip the biography.
The action and the outcome.
One imperative verb gives the model a target. Two verbs usually produce half of each.
Start with write, summarize, review, extract, compare, or plan.
The facts the answer must be built from.
Without source material the model fills gaps with plausible invention. With too much, you pay for tokens it never uses.
Paste the excerpt, the diff, or the notes. Not the whole archive.
Who reads the output.
The same finding becomes a one-line status update for an executive and a three-paragraph explanation for a new engineer.
Name the reader and what they already know.
Rules the answer must respect.
Constraints are where quality lives: no invented metrics, keep the public API stable, quote the source, never guess a date.
Two to four rules, one per line. Include what to avoid.
The shape of the answer.
If you want JSON, a table, or five bullets, say so. Otherwise you get prose and a parsing problem.
Name the structure and the fields or columns it needs.
A hard limit on the answer.
Output tokens are usually the expensive half of a bill, and a cap forces the model to choose what matters.
A number: 150 words, 5 bullets, one paragraph, 400 tokens.
One short input → output pair.
An example is the cheapest way to lock an unusual format or tone that is hard to describe in rules.
Optional. Keep it short; a bad example teaches the wrong thing.
A system prompt is paid for on every request. At 1,000 calls a day, 200 extra tokens of throat-clearing is 6 million tokens a month. On a frontier model that is a real line item; on a cheap model it is noise. The builder shows both numbers while you write, so the trade-off is visible instead of theoretical.
The same logic applies to the answer. A missing length cap lets the model write 600 words when 150 would do. That is not a style problem. It is the expensive half of the bill, and it is the half most prompt guides ignore.
Build for correctness first. Then trim what you can prove is waste. The audit is there to stop you from cutting a constraint that was doing real work.
The template set was inspired by the task-based prompt patterns in Killer ChatGPT Prompts by Guy Hart-Davis (Wiley, 2023): give the model a role and a task, supply the material, name the reader, set the rules, and describe the finished answer. Every template here is original text written for this tool. The book is a useful map of what people actually ask a model to do; the audit and the cost math are ours.
No. It is a structured editor. You fill the fields, it assembles the prompt in your browser, counts the tokens with the o200k_base BPE tokenizer, and prices it against the live model catalog. Nothing is sent to a model or a server.
Chat message mode reads as natural prose, which works better when you paste into a chat window. System prompt mode keeps labeled sections, which is easier for an API to follow consistently and easier for a teammate to edit later. The token counter shows the difference.
It checks whether the prompt contains the eight parts above, then subtracts for filler phrases and vague terms. It is a heuristic, not a grade. A prompt can score 100 and still be wrong for your task. The point is to catch the missing brief before you pay for a retry.
Different providers use different tokenizers. The builder uses o200k_base, the OpenAI tokenizer, for an exact BPE count. Claude, Gemini, Llama, and DeepSeek will land within a few tokens per hundred words. Use the Tokenizer Arena when you need to compare the splits.
The field structure was inspired by the task-based prompt patterns in Killer ChatGPT Prompts by Guy Hart-Davis (Wiley, 2023). Every template here is original text written for this tool, not copied from the book.
Only if the audit still passes. The Trimmer removes filler and whitespace, then rescans the prompt health. If the score drops, it removed something the model needed. Build for correctness first, then cut tokens you can prove are waste.
Token counts use the o200k_base BPE tokenizer via js-tiktoken. Prices are list input rates from the TokenCalc AI catalog and exclude output tokens, caching, batching, retries, and negotiated discounts. The health score is a heuristic, not a guarantee of output quality.
See how OpenAI, Claude, Llama, and DeepSeek tokenizers split the same prompt.
Strip filler, boilerplate, and duplicate context to cut prompt tokens by up to 35%.
Score any prompt for structure, filler, and waste, then trim it with a word-level diff.
We'll email you when model prices change or new tools launch. Unsubscribe anytime.