Text¶
A textual prompt is usually a valid metaprompt:
Hi, LLM! How are you feeling today?
will be expanded to the same string, because it does not contain any MetaPrompt constructs.
Variables¶
Variables should be referenced using this sintax: [:variable_name]. Variable names should match [a-zA-Z_][a-zA-Z0-9_]*.
The syntax for assignments is [:variable_name=any expression].
Optional assignments use ?= instead of = - they update the value only if a variable is unset.
Comments¶
[# Text for the human reader can be written like this.
Comments must be well-formed metaprompt expressions too -
in the future comment parse trees will be used to convey
additional semantic info (e.g. documentation).
Comments are ignored during evaluation.
]
Conditionals¶
[:if the sky is sometimes blue
:then this...
:else that...
]
:if expressions will be expanded at runtime.
First, the following text will be fed to an LLM:
Please determine if the following statement is true.
Do not write any other output, answer just "true" or "false".
The statement: the sky is sometimes blue
The answer will determine the execution branch. If the answer is not literally "true" or "false", an exception will be thrown after a few retries
Meta-prompting¶
LLM says: [$ Hi, LLM! How are you today?]
The [$ prompt will be executed and its output will be inserted
at its position during expansion. This enables powerful techniques
of prompt rewriting:
[$ [$ Improve this LLM prompt: [:prompt]]]
Notice the double nesting of [$ - the improved prompt will be fed back into an LLM.
Chat history¶
Chat history can be preserved by assigning an ID to a prompt: [some_chat_id$ ... ].
Subsequent invocations with the same chat ID within the same module will have a memory of the preceding conversation.
Chat IDs are actually variables that contain an entire chat history (example usage).
Escaping¶
[ and ] must be escaped using \\.
Modules¶
Every .metaprompt file is a module. Conceptually, a module is a function that accepts a number of arguments, runs the executable parts, and returns text.
File imports¶
The following expression will include ./relative-import.metaprompt file (relative to the directory of the file, NOT to the current working dir):
[:use ./relative-import]
Package imports¶
NOT IMPLEMENTED
Passing parameters¶
Unbound variables used in a module are its parameters, that must be provided when calling the module.
Example
Consider a file named hello.metaprompt:
Hello, [:what]! [# `what` is a required parameter ]
[:who?=you]
[# ^ `who` is NOT a required parameter, because it is assigned
before first use. However, optional assignment is used, so
the default value can be overridden from the caller module
]
How are [:who] feeling today?
The hello module can be used from another module:
[:use ./hello :what=world :who=we]
Special variables¶
MODEL switching¶
MODEL variable is used to switch LLM models on the fly (example).
MODEL switching only works before an [:if ... or a [$ ... ] block:
[:MODEL=gpt-4o]
[$ will be run in 4o,
[:MODEL=gpt-3.5-turbo]
[$ but this prompt will run in 3.5-turbo ]
]
Dynamic model selection based on a given task description allows to save costs by avoiding calls to costly models when possible (example).
ROLE switching¶
ROLE is a special variable used to control LLM input "role".
ROLE can be assigned to one of three values:
system: defines the behavior, scope, and context of the LLM.user: represents the individual interacting with the LLM.assistant: represents the LLM itself, responding to the user within the context defined by the system.
See OpenAI docs for more info on roles.
Live STATUS update¶
STATUS variable provides a way to set a status line that is visible in the terminal. Useful to make the user aware of what is happening when no output is being generated:
[:STATUS=running a marathon]
[$ ... long running task ]
[:STATUS=launching the rockets]
[$ ... another long running task ]