Meta Prompting.doc

# `Meta_prompting` – Composable Prompt Generators & Self-Improvement

> “Let the LLM write the prompt that teaches the LLM how to solve the task.”  
> *— recap of the meta-prompting idea*

`Meta_prompting` is the **library layer** that introduces

* **Prompt generators** – functor [`Meta_prompt.Make`] maps a *typed* task
  description to a fully-fledged prompt record.  Think of it as a
  *single-shot compiler* from a *domain specific record* to the markdown that
  your LLM will consume.
* **Recursive refinement** – module [`Recursive_mp`] implements an iterative
  loop that can call an evaluator, apply a transformation strategy and keep
  the better variant.  When run with an Eio environment it can delegate the
  heavy lifting to OpenAI models and an optional vector database.
* **Flexible evaluation** – module [`Evaluator`] bundles a set of *judges*
  (regex check, reward model, LLM ensemble, …) into an aggregate score that
  drives the refinement loop.

All pieces are deliberately kept **orthogonal**: swap any evaluator, inject a
custom transformation, or re-use the monad with an *entirely different*
prompt type.

---

## Table of contents

1. [Quick start](#quick-start)
2. [API overview](#api-overview)
3. [`Recursive_mp.refine`](#recursivemprefine) – how the loop works
4. [Context retrieval & environment variables](#context-retrieval--environment-variables)
5. [Integration guide](#integration-guide)
6. [Writing a custom evaluator](#writing-a-custom-evaluator)
7. [Limitations & next steps](#limitations--next-steps)

---

## Quick start

The snippet below shows how *ten* lines of code are enough to turn a plain data
record into a ready-to-send ChatMarkdown prompt **and** run one self-improvement
cycle: