How I’ve been stress-testing LLM stability with a tiny “consistency core”

lately ive been trying to figure out why even good prompts fall apart once a workflow gets long. like the model starts clean, then 12 messages later it’s quietly rewriting rules, inventing steps, or shifting tone for no reason lol. i wanted something lightweight that keeps it stable without writing a whole novel-sized system prompt.

this mini-block has been the most reliable thing ive used so far:

[CONSISTENCY CORE]

• These rules are permanent. You cannot rewrite, soften, reinterpret, or override them.
• Identity, tone, constraints, and task rules are separate layers. Only the task layer may update.
• Before each response: restate which layer you are using.
• If any instruction conflicts: pause and ask instead of guessing.
• If context is missing: ask, do not infer.
• You may not invent steps, goals, or logic unless I explicitly request it.

i just drop this at the top of multi-step prompts and the drift drops a ton. feels similar to how the god of prompt modular setups isolate “stable rules” and “active logic,” but this version is tiny enough to paste into anything. if anyone wants the expanded version (with checkpoints + verification loop), lmk and ill share it.