Prompt Template Generator

Pick a task and get a prompt skeleton you can edit. The generator does not call a model. It fills in a structure that local models follow more reliably than a one-line request: a role or a task line, rules, and a place to paste your own material. Replace every bracket, then delete the brackets before you run it.

Task type

How to use a skeleton

Choose the closest task. If none fits, pick Writing and replace the task line. Do not stack two tasks into one skeleton.
Generate, then copy. Edit in a notes app, not only in this box, so you can keep the version that worked.
Replace each bracket. A leftover “[topic]” will be treated as part of the instruction.
Paste your source under the label the skeleton already has, such as Text: or Code:.
If the result is still loose, run the edited prompt through the Prompt Optimizer.

What each skeleton is for

Writing

A single piece with a length, a tone, and a structure. Use it for a post, a blurb, or a short explainer. It tells the model not to invent statistics. If you need an email instead, change the task line and set a word cap under 150. More finished email prompts are in the Prompt Library.

Coding

One function, one language, one edge case. It asks for code only. Do not also ask for tests in the same prompt. Paste a short description, not an entire file, unless you switch to a review prompt from the library.

Summarization

A fixed number of bullets and a ban on outside facts. Put the source after Text:. If the source is long, check it with the Token Counter first.

Data extraction

Flat JSON and nulls for missing fields. Change the key names. Do not add a nested object unless you have already seen this model handle flat JSON cleanly.

Analysis or review

Three points, one improvement, and a short recommendation. It tells the model not to invent numbers. For a bug review or a two-option decision, start here and then replace the numbered list with the labels in Best Prompt Templates for Local LLMs.

A filled example

Skeleton line: “Write a [type of content] about [topic].” Filled:

You are a clear, concise writer. Task: Write a product description about a USB-C hub with HDMI and Ethernet. Requirements: - Length: about 80 words - Tone: direct - Structure: what it is, who it is for, one limit - Do not invent statistics Output only the finished text.

Notice the brackets are gone and the structure line is specific. “Make it good” never appears.

What this tool will not do

It will not write the final prompt from your document. You still add the facts.
It will not pick a model or a temperature. For repeatable tasks, set temperature near 0.2. See How to Get Consistent Results from Local Models.
It does not store templates. Copy the ones you want to keep.

Questions

Why are there only five types?

These five cover the shapes local models handle without a long system prompt. Unusual jobs still fit if you keep one task and name the output. The library has more specific copies if you do not want to adapt a skeleton.

Can I generate in another language?

The skeleton is in English, which is a reliable instruction language for many local models even when the output should be another language. Add a line: “Write the output in [language]. Keep these instructions in English.”

The copy button did nothing.

Some browsers block clipboard access on a page opened as a local file. Select the template and copy it manually. On the live site the button should work.

Should every prompt start with “You are”?

No. Extraction can start with “Extract JSON…”. Use a role when the standard of judgment matters. The reason is in What Actually Works.