Templates#

A template can combine a prompt, system prompt, model, default model options, schema, and fragments into a single reusable unit.

Only one template can be used at a time. To compose multiple shorter pieces of prompts together consider using fragments instead.

Getting started with –save#

The easiest way to create a template is using the --save template_name option.

Here’s how to create a template for summarizing text:

llm '$input - summarize this' --save summarize

Put $input where you would like the user’s input to be inserted. If you omit this their input will be added to the end of your regular prompt:

llm 'Summarize the following: ' --save summarize

You can also create templates using system prompts:

llm --system 'Summarize this' --save summarize

You can set the default model for a template using --model:

llm --system 'Summarize this' --model gpt-4o --save summarize

You can also save default options:

llm --system 'Speak in French' -o temperature 1.8 --save wild-french

If you want to include a literal $ sign in your prompt, use $$ instead:

llm --system 'Estimate the cost in $$ of this: $input' --save estimate

Add --schema to bake a schema into your template:

llm --schema dog.schema.json 'invent a dog' --save dog

If you add --extract the setting to extract the first fenced code block will be persisted in the template.

llm --system 'write a Python function' --extract --save python-function
llm -t python-function 'reverse a string'

In each of these cases the template will be saved in YAML format in a dedicated directory on disk.

Using a template#

You can execute a named template using the -t/--template option:

curl -s https://example.com/ | llm -t summarize

This can be combined with the -m option to specify a different model:

curl -s https://llm.datasette.io/en/latest/ | \
  llm -t summarize -m gpt-3.5-turbo-16k

Templates can also be specified as full URLs to YAML files:

llm -t https://raw.githubusercontent.com/simonw/llm-templates/refs/heads/main/python-app.yaml \
  'Python app to pick a random line from a file'

Or as a direct path to a YAML file on disk:

llm -t path/to/template.yaml 'extra prompt here'

Listing available templates#

This command lists all available templates:

llm templates

The output looks something like this:

cmd        : system: reply with macos terminal commands only, no extra information
glados     : system: You are GlaDOS prompt: Summarize this:

Templates as YAML files#

Templates are stored as YAML files on disk.

You can edit (or create) a YAML file for a template using the llm templates edit command:

llm templates edit summarize

This will open the system default editor.

Tip

You can control which editor will be used here using the EDITOR environment variable - for example, to use VS Code:

export EDITOR="code -w"

Add that to your ~/.zshrc or ~/.bashrc file depending on which shell you use (zsh is the default on macOS since macOS Catalina in 2019).

You can create or edit template files directly in the templates directory. The location of this directory is shown by the llm templates path command:

llm templates path

Example output:

/Users/simon/Library/Application Support/io.datasette.llm/templates

A basic YAML template looks like this:

prompt: 'Summarize this: $input'

Or use YAML multi-line strings for longer inputs. I created this using llm templates edit steampunk:

prompt: >
    Summarize the following text.

    Insert frequent satirical steampunk-themed illustrative anecdotes.
    Really go wild with that.

    Text to summarize: $input

The prompt: > causes the following indented text to be treated as a single string, with newlines collapsed to spaces. Use prompt: | to preserve newlines.

Running that with llm -t steampunk against GPT-4o (via strip-tags to remove HTML tags from the input and minify whitespace):

curl -s 'https://til.simonwillison.net/macos/imovie-slides-and-audio' | \
  strip-tags -m | llm -t steampunk -m gpt-4o

Output:

In a fantastical steampunk world, Simon Willison decided to merge an old MP3 recording with slides from the talk using iMovie. After exporting the slides as images and importing them into iMovie, he had to disable the default Ken Burns effect using the “Crop” tool. Then, Simon manually synchronized the audio by adjusting the duration of each image. Finally, he published the masterpiece to YouTube, with the whimsical magic of steampunk-infused illustrations leaving his viewers in awe.

System prompts#

When working with models that support system prompts you can set a system prompt using a system: key like so:

system: Summarize this

If you specify only a system prompt you don’t need to use the $input variable - llm will use the user’s input as the whole of the regular prompt, which will then be processed using the instructions set in that system prompt.

You can combine system and regular prompts like so:

system: You speak like an excitable Victorian adventurer
prompt: 'Summarize this: $input'

Fragments#

Templates can reference Fragments using the fragments: and system_fragments: keys. These should be a list of fragment URLs, filepaths or hashes:

fragments:
- https://example.com/robots.txt
- /path/to/file.txt
- 993fd38d898d2b59fd2d16c811da5bdac658faa34f0f4d411edde7c17ebb0680
system_fragments:
- https://example.com/systm-prompt.txt

Options#

Default options can be set using the options: key:

name: wild-french
system: Speak in French
options:
  temperature: 1.8

Schemas#

Use the schema_object: key to embed a JSON schema (as YAML) in your template. The easiest way to create these is with the llm --schema ... --save name-of-template command - the result should look something like this:

name: dogs
schema_object:
    properties:
        dogs:
            items:
                properties:
                    bio:
                        type: string
                    name:
                        type: string
                type: object
            type: array
    type: object

Additional template variables#

Templates that work against the user’s normal prompt input (content that is either piped to the tool via standard input or passed as a command-line argument) can use the $input variable.

You can use additional named variables. These will then need to be provided using the -p/--param option when executing the template.

Here’s an example YAML template called recipe, which you can create using llm templates edit recipe:

prompt: |
    Suggest a recipe using ingredients: $ingredients

    It should be based on cuisine from this country: $country

This can be executed like so:

llm -t recipe -p ingredients 'sausages, milk' -p country Germany

My output started like this:

Recipe: German Sausage and Potato Soup

Ingredients:

  • 4 German sausages

  • 2 cups whole milk

This example combines input piped to the tool with additional parameters. Call this summarize:

system: Summarize this text in the voice of $voice

Then to run it:

curl -s 'https://til.simonwillison.net/macos/imovie-slides-and-audio' | \
  strip-tags -m | llm -t summarize -p voice GlaDOS

I got this:

My previous test subject seemed to have learned something new about iMovie. They exported keynote slides as individual images […] Quite impressive for a human.

Specifying default parameters#

When creating a template using the --save option you can pass -p name value to store the default values for parameters:

llm --system 'Summarize this text in the voice of $voice' \
  --model gpt-4o -p voice GlaDOS --save summarize

You can specify default values for parameters in the YAML using the defaults: key.

system: Summarize this text in the voice of $voice
defaults:
  voice: GlaDOS

When running without -p it will choose the default:

curl -s 'https://til.simonwillison.net/macos/imovie-slides-and-audio' | \
  strip-tags -m | llm -t summarize

But you can override the defaults with -p:

curl -s 'https://til.simonwillison.net/macos/imovie-slides-and-audio' | \
  strip-tags -m | llm -t summarize -p voice Yoda

I got this:

Text, summarize in Yoda’s voice, I will: “Hmm, young padawan. Summary of this text, you seek. Hmmm. …

Configuring code extraction#

To configure the extract first fenced code block setting for the template, add this:

extract: true

Setting a default model for a template#

Templates executed using llm -t template-name will execute using the default model that the user has configured for the tool - or gpt-3.5-turbo if they have not configured their own default.

You can specify a new default model for a template using the model: key in the associated YAML. Here’s a template called roast:

model: gpt-4o
system: roast the user at every possible opportunity, be succinct

Example:

llm -t roast 'How are you today?'

I’m doing great but with your boring questions, I must admit, I’ve seen more life in a cemetery.

Template loaders from plugins#

LLM plugins can register prefixes that can be used to load templates from external sources.

llm-templates-github is an example which adds a gh: prefix which can be used to load templates from GitHub.

You can install that plugin like this:

llm install llm-templates-github

Use the llm templates loaders command to see details of the registered loaders.

llm templates loaders

Output:

gh:
  Load a template from GitHub or local cache if available

  Format: username/repo/template_name (without the .yaml extension)
    or username/template_name which means username/llm-templates/template_name

Then you can then use it like this:

curl -sL 'https://llm.datasette.io/' | llm -t gh:simonw/summarize

The -sL flags to curl are used to follow redirects and suppress progress meters.

This command will fetch the content of the LLM index page and feed it to the template defined by summarize.yaml in the simonw/llm-templates GitHub repository.

If two template loader plugins attempt to register the same prefix one of them will have _1 added to the end of their prefix. Use llm templates loaders to check if this has occurred.