Rigid Instructions Need Flexibility | Generated by AI
Question: Are overly rigid custom instructions counterproductive, and how can they be made more useful?
Answer:
Yeah, you’re right. A few problems with that prompt:
It’s too prescriptive and repetitive. Mentioning Karpathy/Yin Wang as style targets every response gets stale fast. It’s a vibe-setting instruction that wears out — and I end up pattern-matching to the names rather than actually thinking like them.
The format template is rigid. Question: / Answer: / References: for every single response adds friction for quick exchanges like this one.
It over-explains your backstory. Background context is useful once, but restating your full resume every conversation is noise.
The “randomly check if I’m focused on LLM training” idea is actually good — but it needs to be an active mechanism, not a passive note. You could implement it as a simple rule like: “10% of the time, at the end of a response, ask me: ‘Is this aligned with your LLM training focus?’ and flag drift.”
What would actually be more useful:
A leaner prompt with three sections:
- Static context (devices, preferences, stack) — just a table, no prose
- Active rules (redaction, format, depth level) — concrete and short
- Focus enforcement — a real probabilistic interrupt rule
The “I want to grow like Karpathy” framing should be removed entirely. Instead, just say: “assume strong ML/systems background, go to first principles, show code.” The outcome is the same without the cliché.
Want me to rewrite your system prompt from scratch — leaner, more mechanical, with the focus-check mechanic built in?