Problem: Teams often develop AI assistants ad hoc, copy prompts from ChatGPT, or generate hastily written instructions that behave unpredictably in production. The result are inconsistent outputs, bias in evaluative tasks, and a lack of traceability when formats deviate from expectations. Companies running multiple assistants in parallel also lack a unified architecture framework that documents, from the discovery phase through to model selection, why specific design decisions were made.

Solution: A specialized meta-assistant takes over the methodical development of production-ready system prompts using a three-phase process. Phase one is discovery: even when requirements appear clear, the assistant asks follow-up questions about the target audience, input format, style rules, and known failure modes before it starts building anything. Phase two is structured prompt creation based on an enterprise framework with mandatory sections such as role, context, task, output format, and rules, as well as optional sections such as evaluation criteria or negative examples. Phase three delivers a complete package consisting of an assistant name, one-line description, full system prompt in a code block, design decision table, model recommendation based on task complexity, and a self-checklist covering all relevant quality criteria.

Use Cases: A customer service team could build an assistant that classifies email inquiries by urgency and suggests prewritten response modules — the discovery phase would automatically ask about tone-of-voice guidelines and escalation criteria. A sales team could have meeting notes converted into structured follow-up tasks, with negative examples showing which vague formulations should be avoided. A controlling team could translate monthly Excel reports into management summaries, with the model recommendation weighing fast processing with Gemini Flash against analytical depth with Claude Sonnet. A marketing team could categorize social media comments by sentiment, with anti-bias rules ensuring that ironic or culture-specific expressions are interpreted correctly.

Explanatory Approach: Traditional prompt development often fails because LLMs make implicit assumptions that lead to deviations in production. The meta-assistant enforces explicitness through discovery questions and documents every design decision — for example, why an optional section for negative examples was chosen or which guardrails were implemented against common LLM errors. The framework deliberately separates formatting rules from style rules and prevents the role definition from introducing bias in evaluative tasks. The self-checklist systematically verifies whether the input format is clearly described, whether anti-bias measures have been implemented, and whether all structural requirements are met. The model recommendation matrix makes transparent why Claude Opus is recommended for meta-tasks, while GPT Mini is sufficient for simple structuring tasks.

Conclusion: The meta-assistant professionalizes assistant development through methodological rigor instead of trial and error. Teams gain a reusable architecture pattern that documents how production-ready systems are created, from the first discovery question to the final checklist. The investment in structured prompt development pays off as soon as assistants are no longer used for local experimentation, but deployed across the enterprise for critical workflows.

Interested in the full system prompt for the meta-assistant? Feel free to contact daniel@blackmountain.io.