Hi everyone,

One issue I encounter often when I write complex prompts or skills is the agent interpreting my words the wrong way, or deciding to batch or skip steps.

Problem: Human language is inherently ambiguous and when we add that to LLMs tendency to hallucinate and take shortcuts, we get very unpredictable results.

Idea: what if we wrote prompts in a programming language? this should force the agent to follow it step by step, right?

I chose Typescript as the language and built mantiq, a cli and a skill that let's you write prompts and skills using Typescript, some features/benefits are:

• Having full control over the execution of the prompt.
• The ability to implement determistic logic as Typescript functions that will be called as tools by the agent.
• The ability to tell the agent what to think about or which actions to take at any step of the program.

I am still experimenting with this tool and would love to have your feedbacks on it: https://github.com/webNeat/mantiq