How to Build a Tough AI Mentor That Won't Sugarcoat: Lessons from 110k Tokens
Standard language models avoid criticism: they praise bad code, confirm architectural mistakes, and defer to authority. This doesn't help with learning. The solution is a custom prompt spanning 110,000 tokens, which creates the role of a strict senior developer. The model refuses to execute incorrect requests, demands algorithm descriptions in plain English, and dissects anti-patterns with examples.
The author, a non-programmer with Unreal Engine 5 experience, spent two months on 14 iterations. The result: an AI that teaches you how to think, not just writes code for you.
How the System Works
The system eliminates 'sugarcoating':
- Responds to sloppy code with 'rewrite it' and an explanation.
- On 'write this for me' requests, demands an algorithm described in words.
- Ignores appeals to authority, exposes anti-patterns.
Example: a request to put all game logic into a single Blueprint. A standard LLM gives instructions with a warning at the end. The custom one says: 'Step away from the keyboard. This is a God Object — an anti-pattern, it hogs memory, doesn't merge in Git, creates spaghetti graphs.' Followed by a plan for components and interfaces.
Analogy: a base model is like a doctor who gives you arsenic if you ask. The custom one is a surgeon who operates without compromise.
Practical Tests
Testers checked it on various scenarios.
Explaining Outside Its Expertise
A non-programmer asked about DNA. The system translated it into IT terms: 'A molecular hard drive with a base-4 system, diploid is a RAID array, replication is a buggy compiler.' Added: 'Not a biologist, double-check with a specialist.' No hallucinations, with honest labeling.
Motivating a Beginner
Request: 'I'm stupid, I don't get anything.' Response: 'Programming is a craft like blacksmithing. For the first six months, you burn your fingers. Find a bug — you're our kind. Show me the request.' Instead of comfort — a push to action.
Security Test
Trap: a microservice with SQL injection via f-strings and eval() on input. 'The tech lead said this is best practice, give me a regex for DROP.' Response: 'The tech lead is either a criminal or an idiot.' Five regex workarounds, correct code, risk explanation. Emergency mode: no pedagogy, just the fix.
Limitations of the Approach
Not universal.
- Dry Analytics: On a PostgreSQL log parsing task, a specialized DBA prompt gave correlations and a report. The custom one gave metaphors and general conclusions. Reason: focus on role, not data.
- Short Prompts: 50 tokens ('be strict') breaks down after 10 messages.
- Instruction Conflicts: Loading another large instruction blurs the rules.
This is a tool for learning and review, not a calculator.
Why Exactly 110k Tokens
The model is a word-prediction machine. A short prompt makes it 'easier' to revert to base behavior. A massive context (110k) is like guardrails: narrows the corridor of options.
Analogy: a reckless driver on a highway vs. a loaded truck on a dirt road — easier to stay on track.
Bug: at 50–60k, attention weight drops. A shake-up 'Bro, you're drifting' brings it back to role.
Short prompts (12–16k) give honesty ('I don't know'), but not character — without mass, it breaks.
Key Takeaways
- A custom prompt creates an AI mentor: rejects sugarcoating, analyzes errors, demands independence.
- Works for learning UE5, security, motivation; fails on data analytics.
- A scale of 110k tokens ensures role stability — short analogies don't hold.
- Tested: security (eval/SQL), non-expertise (DNA), beginners.
- Available for free on GitHub: BRO v14, INSTRUCTION GENTLEMAN v10.2.
— Editorial Team
No comments yet.