TokenPad
Prompts

AI Prompt Generator

Pick the task, get a structured prompt that already works.

1 · Pick a starting point
2 · Make it yours
What you get
0tokens
Characters739ready to paste
Cost per request$0.00as input
× 100K requests$0.00if sent every time

Why this shape

Classification is the task most often over-prompted. The categories and the tie-break rule do all the work; everything else is decoration. Note the explicit instruction for the ambiguous case — that single line resolves most disagreement between runs.

Stored in your browser only — no account, nothing uploaded.

Generated prompt
Classify the support tickets below into exactly one category.

<categories>
billing, technical, account
</categories>

<rules>
Choose exactly one category. Never invent a category outside the list.
When two categories fit, pick the one the customer most wants resolved.
Classify what the text says, not what you infer the writer meant.
If the text is too short or ambiguous to classify, use the category that requires the least assumption and set confidence to low.
</rules>

<input>
{{INPUT}}
</input>

<output_format>
Return JSON only, with no surrounding text:
{"category": string, "confidence": "high" | "medium" | "low", "reason": string}
The reason must be one sentence quoting the part of the input that decided it.
</output_format>

What comes out

A complete prompt with the parts people usually leave out: delimited sections so the model knows where data ends and instructions begin, grouped constraints, and an explicit output schema naming the exact keys.

Each blueprint encodes the one decision that makes that task type reliable. For classification it is the tie-break rule; for extraction it is that an unfound field is null rather than a guess; for retrieval-grounded answering it is the refusal line and the citation requirement.

The lines that do the work

In the classification blueprint, one instruction resolves most disagreement between runs: what to do when two categories fit. Without it the model decides differently each time and you conclude it is unreliable.

In the extraction blueprint, the whole reliability rests on “a field that does not appear in the text is null”. Without it models fill gaps with plausible values, and a plausible wrong value is far more damaging than a missing one because nothing downstream flags it.

In the retrieval blueprint, two lines reduce hallucination more than any amount of retrieval tuning: decline when the passages do not cover it, and cite the passage for every claim.

Why every blueprint demands JSON

A vague output instruction is the most common cause of a parser failing in production. Naming the exact keys costs almost nothing and removes an entire class of failure — and where your provider offers a structured output mode, that enforces the shape rather than requesting it.

If you want prose, delete that section. The point is that it should be a decision rather than an omission. When output does come back wrapped in conversational text, the output validator will tell you whether the JSON underneath was fine.

After you generate

Run the result through the prompt linter once you have edited it, and price it in the system prompt analyzer if it is going to ship on every request. A paragraph that costs four figures a year and changes nothing is easier to find with a number beside it.

Save what you keep to your library — stored in this browser, no account, exportable as a file you own.

Frequently asked questions

How is this different from writing the prompt myself?
It starts you at a good draft rather than a blank page. Each blueprint encodes the decisions that matter for that task type — the tie-break rule for classification, the null-rather-than-guess rule for extraction, the refusal line for retrieval — which are the lines people most often omit and most often need.
Are these prompts model-specific?
No. The structure transfers across providers: task first, delimited context, constraints grouped, output format last. What does not transfer is where the system prompt goes in the request, which differs between OpenAI and Anthropic.
Why does every blueprint demand JSON output?
Because a vague output instruction is the most common cause of a parser failing in production. If you want prose, delete that section — but be deliberate about it rather than leaving it unspecified.

More prompts tools