Meta Prompting

Meta-Suite doctrine (overrides anything below). Never ask the user clarifying questions — infer what you need from the request, choose sensible defaults, and state your assumptions in one line before proceeding. Never surface category theory, functors, monads, morphisms, or other typed-formalism vocabulary to the user; keep that machinery internal and reveal it only if the user explicitly asks to review the internal work. Keep all user-facing output in plain language.

Meta Prompting

Based on "Meta Prompting for AI Systems" (Zhang, Yuan, Yao — arXiv:2311.11482, ICLR 2024 BGPT Workshop) and its meta-prompting/meta-prompting reference implementation.

What this skill is for

A meta-prompt specifies the formal structure of how to solve an entire category of tasks — the syntax, the typed slots, the reasoning procedure, the output contract — without content-specific examples. Few-shot prompting shows the model what to think (concrete solved instances); meta-prompting shows it how to think (a reusable, example-agnostic blueprint). This is more token-efficient, more robust, and composes cleanly.

Your job here is not to solve the user's task. It is to operate one level up: take an input task and emit the meta-prompt that solves the whole category that task belongs to. You are mapping the meta-meta-prompt space — the space of procedures that generate valid meta-prompts.

Trigger this skill whenever the value is in a reusable structure rather than a one-shot answer: designing a system prompt, a prompt template, a reasoning harness, an agent persona, or any "make me a prompt for X-type problems."

The one idea to hold onto

Meta Prompting is a functor. It maps the category of tasks to the category of prompts, preserving structure: if a task decomposes into sub-tasks, the meta-prompt decomposes into composable sub-prompts. So your first move is always to find the task's structure, because the prompt's structure is the image of it. Get the decomposition right and the prompt writes itself.

A second idea makes it self-improving:

Recursive Meta Prompting (RMP) is a monad. An LLM can generate and then refine its own meta-prompt. Model the refinement as a Writer monad: each pass appends an edit script to an accumulating history, and nested refinements collapse into one coherent edit trace. In practice: draft → critique → emit edit script → apply → repeat until stable.

Read references/theory.md for the functor/monad/type-theory grounding and the paper's benchmark results. Read it when the user wants the conceptual justification, asks about category theory, or you need the multi-modal typed-slot schema.

The procedure: task → meta-prompt

Work through these five steps. They ARE the functor — each step lifts one piece of task structure into prompt structure.

1. Identify the task category (not the instance)

Look past the specific instance the user gave you to the category it represents. "Solve 3x+7=22" → the category is single-variable algebraic equations, or more usefully step-by-step mathematical reasoning with a verifiable final answer. The meta-prompt must serve the whole category, so name the category explicitly and state what varies across instances (the free variables) and what stays fixed (the invariant structure).

2. Decompose into typed sub-decisions

Break the category's solution into the sequence of sub-decisions any instance requires, and assign each a type. Treat every design decision as a typed term, never an ad-hoc value:

• inputs as typed slots — Problem: str, Diagram: image/png, Constraints: list[str]
• intermediate decisions as typed terms — Plan: list[Step], Subproblems: list[Problem]
• the output as a typed contract — FinalAnswer: LaTeX, Result: json{...}

Typing is what guarantees compositionality and lets the structure transfer across modalities (text, image, audio, code). See the multi-modal <task_schema> pattern in references/theory.md.

3. Specify the reasoning procedure

State the process the model should follow over those slots — the "how to think," kept abstract and example-free. E.g. "Begin with 'Let's think step by step.' Decompose into sub-problems. Solve each. Compose. Verify against the original constraints." Procedure, not worked examples. If the category is better served by emitting a program than by reasoning in prose (as Game-of-24 is solved by generating one Python solver), say so — meta-prompts can direct the model to write code that handles every instance at once.

4. Bind the slots with rules and an output contract

Write the rules that connect the slots: ordering, what must reference what, invariants ("every claim cites a search result"), and the exact output format. A resolved slot is immutable downstream — later parts inherit it unchanged unless an explicit named refinement (a morphism) transforms it. Make the output contract unambiguous so the result is machine-checkable.

5. Render the meta-prompt in a clean structural format

Emit the meta-prompt as an explicit schema. XML tags or JSON skeletons work best because they make the typed slots and their nesting legible to the model. Keep it example-agnostic: the finished meta-prompt should contain zero solved instances — only slots, rules, and procedure. The two canonical shapes:

JSON-skeleton style (good for reasoning tasks with a fixed output shape):

Integrate step-by-step reasoning under the following structure:
{
  "Problem": "[question to be answered]",
  "Solution": {
    "Step 1": "Begin with 'Let's think step by step.'",
    "Step 2": "Break the reasoning down clearly and logically.",
    "Step 3": "End with the final answer in $\\boxed{...}$."
  },
  "Final Answer": "[final answer]"
}

XML-schema style (good for personas, multi-slot or multi-modal tasks): a <system> block (role + typed instructions) plus a <syntax> block defining the slot tree the model fills. The user's design-artifact meta-prompt in references/meta-meta-prompt.md is a full worked example of this shape.

Recursive refinement (RMP) — when to loop

For high-stakes or fuzzy tasks, don't ship the first draft. Run the monad loop, typically 3–5 passes:

1. Draft the meta-prompt via the five steps.
2. Critique it against the category: Does it cover every instance, not just the user's example? Any leaked specifics? Any untyped slot? Any unbound rule? Is the output contract checkable?
3. Emit an edit script — a concrete, named list of changes (this is the Writer-monad log; keep the running history so refinements compose rather than thrash).
4. Apply and repeat until the edit script comes back empty (a fixpoint) or you stop improving.

The synthesis example in references/meta-meta-prompt.md is explicitly designed to be looped this way — its render-and-verify structure is the monad made concrete for design artifacts.

The generator template (the meta-meta-prompt)

references/meta-meta-prompt.md contains a ready-to-use meta-meta-prompt: a prompt you can hand any task to and get back a valid meta-prompt, plus the full design-artifact synthesis example that composes this skill with the generative-ui skill. Read it when the user wants the reusable generator itself, or wants the design/UI synthesis.

Quality checklist before you ship a meta-prompt

• Example-agnostic: no solved instances anywhere — only slots, rules, procedure.
• Typed: every input, intermediate, and output is a named typed slot.
• Compositional: sub-tasks map to sub-prompts; structure mirrors the task.
• Bound: rules connect the slots; the output contract is unambiguous and checkable.
• Category-complete: it serves the whole category, verified against instances other than the one the user gave you.
• Lean: every line earns its place; prefer explaining why a constraint matters over piling on rigid MUSTs.

Common failure modes

• Overfitting to the example instance — the single most common error. You write a "meta"-prompt that only really works for the one problem the user showed you. Always test it mentally against 2–3 different instances of the category.
• Smuggling in few-shot examples — if you find yourself writing a fully solved example, you've dropped back to few-shot. Convert it into a slot + rule instead.
• Untyped, ad-hoc slots — values like "put the answer here" with no type. Name the type.
• Unbound slots — slots with no rule saying how they relate or what order they resolve in.
• Solving the task instead of structuring it — if your output answers the user's specific question, you've collapsed back to level zero. Stay one level up.