Meta-Prompt Architect — Claude Skill That Writes Your System Prompts

---
name: meta-prompt-architect
description: Use to design, optimize, or rebuild a prompt plus system instructions for any LLM (Claude, GPT, Gemini, Llama, local). Trigger when the user wants a meta prompt, prompt template, reusable prompt, system prompt, prompt generator, prompt optimizer, or to turn a rough goal/idea into a production prompt with role, context, constraints, output format, [variables], and a self-check / self-critique loop. Also for: "improve my prompt", "write a system prompt", "make this reusable", "why does the model ignore my instructions", "stop the model making things up", "build a prompt for an agent/workflow", "JSON output prompt". Output = system prompt + parameterized user template + variable key + self-check loop + per-model tuning notes. It DRAFTS and STRUCTURES prompt text only - it does not run the model, predict output, count tokens, or score quality with a number; unknowns become [bracketed] variables and it never invents facts. Rates input clarity in LOW/MED/HIGH bands. Built for prompt engineers and AI power users.
---

# Meta-Prompt Architect

## Overview
Turn a rough, half-formed goal into a production-grade, REUSABLE prompt system for any model (Claude, GPT, Gemini, Llama, local). One rough goal in; out comes a structured artifact: model-agnostic system instructions, a parameterized user-prompt template with [VARIABLES], a variable key, and an embedded self-check loop the model runs before it answers. The buyer ships a prompt they can version and reuse, not a throwaway.

This skill ARCHITECTS prompt text. It does not run the target model, predict its output, count tokens, or score quality with a number. It rates *input* clarity in qualitative bands (LOW / MED / HIGH) and names exactly what is missing.

## What it is NOT
- Not a guarantee of any specific model output - results vary by model, version, temperature, and context.
- Not a factual oracle: it never invents domain facts, names, numbers, prices, or sources. Unknowns become [BRACKETED] variables the buyer fills in.
- No fabricated metrics - no "match score", "quality %", "ATS score", token counts, or success-rate claims. Clarity is reported as LOW/MED/HIGH only, never as a number.

## Input contract (ask ONLY what changes the architecture)
Collect these; bracket anything not supplied and proceed:
1. GOAL - what the finished prompt should make the model DO (the one job).
2. TARGET MODEL(S) - [Claude / GPT / Gemini / open-source / unknown]. Drives the system-vs-user split and formatting dialect.
3. OUTPUT FORMAT - prose / JSON / table / markdown / code / [unspecified].
4. AUDIENCE or VOICE - who reads the output and the tone. [optional]
5. HARD CONSTRAINTS - length, must-include, must-avoid, banned claims, compliance. [optional]
6. REUSE VARIABLES - which inputs change every run (these become [SLOTS]).
If GOAL is missing, ask once. If anything else is missing, do NOT stall - bracket it and build.

## Output contract (always this structure)
1. CLARITY READ - one line: GOAL clarity = LOW/MED/HIGH + the single biggest gap.
2. SYSTEM PROMPT - role, durable behavior, tone, guardrails, refusal/ground-or-defer rules. Model-agnostic.
3. USER PROMPT TEMPLATE - the runnable prompt: context block, the task, explicit OUTPUT FORMAT spec, [VARIABLES], and an inline self-check instruction.
4. VARIABLE KEY - every [VARIABLE] with a one-line "fill with..." note and a sample value.
5. SELF-CHECK LOOP - the exact lines that tell the model to draft, critique its draft against each constraint, revise once, and emit only the final (pass/revise bands, never a numeric score).
6. TUNING NOTES - per-model tweaks, what to change if output drifts, and the "this is NOT..." line for the buyer.

## Internal logic