Building an AI Tutor with Prompt Engineering: Inside Mr. Ranedeer’s 5,000-Token System Prompt
Hook
A single prompt system convinced GPT-4 to become a PhD-level tutor that remembers your learning style, adjusts explanation depth on command, and generates custom tests—all without fine-tuning a single parameter.
Context
When GPT-4 launched, most developers building educational AI faced a choice: invest in fine-tuning custom models, or accept generic ChatGPT responses that treated every learner identically. The problem wasn’t capability—GPT-4 could explain quantum physics or elementary arithmetic—but adaptability. A middle schooler and a doctoral candidate asking about partial fractions needed fundamentally different pedagogical approaches.
Mr. Ranedeer AI Tutor emerged as a third path: sophisticated prompt engineering that transforms a base language model into a stateful, personalized teaching system. Created by JushBJJ and now discontinued after reaching 29,667 stars, it demonstrated that carefully architected prompts could deliver experiences previously requiring custom model training. The project became a landmark in prompt engineering, showing that the right system instructions could create persistent personas, multi-dimensional configuration systems, and command-driven interfaces—all within ChatGPT’s existing infrastructure.
Technical Insight
The core innovation in Mr. Ranedeer is its meta-prompt architecture: a structured configuration layer that precedes every tutoring interaction. Rather than asking GPT-4 to ‘be a tutor,’ it establishes a complete behavioral framework with 10 depth levels (elementary through PhD), 6 learning styles (visual, verbal, active, intuitive, reflective, global), 5 communication formats (storybook, textbook, layman, story telling, Socratic), 5 tone styles (encouraging, neutral, informative, friendly, humorous), and 5 reasoning frameworks (deductive, inductive, abductive, analogical, causal). This multi-dimensional matrix creates numerous possible tutor configurations from a single prompt.
The system implements pseudo-statefulness through explicit rule definitions that persist across conversation turns. When you execute /config, you’re triggering a prompt pattern that tells GPT-4 to regenerate its understanding of your preferences and format subsequent responses accordingly. The prompt defines command handlers including /config (guide through configuration), /plan (create lesson plan), /start (start the lesson), /test (assess knowledge), and /language (change tutor language).
The depth personalization provides concrete differentiation across 10 levels, from Elementary (Grade 1-6) through PhD. This creates measurable differences in how content is presented—ask about partial fractions at elementary level versus doctoral level and you’ll receive dramatically different explanations suited to each knowledge tier.
The prompt also implements a format templating system. When configured for Socratic reasoning and encouraging tone, GPT-4 appears to structure responses as guided questions with positive reinforcement. The learning style modifier further adjusts output—visual learners may receive more spatial analogies, while active learners get more experimental prompts.
The entire system lives in the initial context window. The v2.7 Reboot version consumes 5,376 + 200 + 247 tokens before any actual teaching begins. This front-loaded architecture means configuration can survive across multiple exchanges until context limits force truncation, though the trade-off sacrifices initial tokens for persistent behavioral shaping.
Cross-LLM portability proved surprisingly robust. While optimized for GPT-4, users reported success with Claude-100k according to project documentation. The underlying principle—detailed behavioral specification through structured instructions—appears to transfer across transformer architectures, though output quality varies with base model capability. GPT-3.5 technically executes the prompt but the README notes it ‘will not be as effective and concise as GPT-4.‘
Gotcha
The elephant in the room: Mr. Ranedeer is officially discontinued. The README leads with ‘DISCONTINUED’ in all caps, meaning no security patches, no compatibility updates as OpenAI modifies GPT-4’s behavior, and no new features. This isn’t abandonware you can fork and maintain—the value proposition depends entirely on GPT-4’s continued responsiveness to these specific prompt patterns, which OpenAI may inadvertently break in future model updates.
Token economics present a practical consideration. The v2.7 Reboot initialization consumes 5,376 + 200 + 247 tokens before any teaching occurs. For API users, this represents a per-session cost before any actual content. The ChatGPT Plus web interface sidesteps per-token billing but requires the subscription, and the prompt explicitly warns against using GPT-4 with plugins or without Code Interpreter as of v2.7—limiting your deployment options.
Context window limitations create boundaries on lesson complexity. After the prompt initialization, you’re working with remaining tokens in GPT-4’s context window. The system has no external memory—once context fills, early configuration may get truncated and behavior could drift.
The personalization depth is simultaneously the strength and weakness. With multiple configuration dimensions, finding your optimal setup requires experimentation. The repository includes a ‘Config Wizard’ GPT to help, but that’s another interface to learn, another conversation to initiate. For learners who just want to ask questions, the configuration overhead may exceed the benefit.
Verdict
Use Mr. Ranedeer if you’re a self-directed learner with ChatGPT Plus who wants maximum control over AI tutoring interactions and you’re comfortable with a discontinued tool that may degrade unpredictably. It appears best suited for advanced students (undergraduate+) exploring technical subjects where depth personalization matters—think abstract algebra, advanced algorithms, or graduate-level physics where generic explanations fall flat. It’s also invaluable as a reference implementation if you’re building prompt-based systems yourself; the structured configuration patterns and command handling techniques may transfer to other LLM applications. Skip it if you need institutional support or long-term reliability (it’s unmaintained), if you want plug-and-play simplicity (configuration is extensive), or if you’re primarily teaching elementary/middle school students with straightforward questions where the configuration overhead may outweigh benefits. The project’s legacy is less about being a production tutoring tool and more about proving that prompt engineering could achieve sophistication previously requiring fine-tuning—a lesson that remains relevant even as the specific implementation fades.
