By skyforbes I’ve been working on a tutoring agent that runs three internal modes (lesson delivery, guided practice, and user-uploaded question review). It uses guardrails like:
- a strict four-step reasoning sequence,
- no early answer reveals,
- a multi-tier miss-logic system,
- a required intake phase,
- and a protected “static text” layer that must never be paraphrased or altered.
The whole thing runs on text only—no functions, no tools—and it holds state for long sessions.
I’m not planning to post the prompt itself, but I’m absolutely open to critiques of the approach, structure, or architecture. I’d really like feedback on:
- Guardrail stability: how to keep a large rule set from drifting 15–20 turns in.
- Mode-switching: ideal ways to route between modes without leaking internal logic.
- “Protected text” handling: making the model respect verbatim modules without summarizing or synthesizing them.
- Error handling: best practices for internal logging without revealing system details to the user.
- Long-session resilience: strategies for keeping tone and behavior consistent over 100+ turns.
If you’ve built similarly complex, rule-heavy agents, I’d love to compare notes and hear what you’d do differently.
https://chatgpt.com/g/g-691ac322e3408191970bd989a69b3003-chatty-the-sat-reading-tutor