RES, my personal safeguard against AI

The author modifies the system prompt of a conversational AI—likely a ChatGPT‑style model—by inserting a rule that each response must begin with the word “RES.” This tiny constant acts as a quick sanity check: if the token disappears, the user knows the model has ignored a prior constraint and must reread the answer. The technique mirrors the “canary token” concept from cybersecurity, where a hidden string alerts defenders to prompt‑injection attacks. Here the token is visible and user‑verified, turning a simple string into a manual monitoring tool for instruction fidelity.

The trick surfaces amid a wave of user‑level prompt engineering as developers grapple with “hallucinations” and instruction drift in large language models. Platforms such as OpenAI now expose custom instructions and system prompts, encouraging users to shape model behavior without writing code. By borrowing a security metaphor, the author demonstrates how end‑users can create low‑cost safeguards that complement—rather than replace—formal safety layers. This DIY approach reflects a broader trend: practitioners are experimenting with lightweight, repeatable patterns to keep AI outputs aligned, especially in long‑running chats where context windows grow and subtle deviations become harder to spot.

While the “RES” canary is inexpensive and easy to deploy, it offers only a binary signal and depends on the model consistently obeying the rule. Future model updates could ignore or overwrite custom tokens, and malicious actors might deliberately trigger or suppress the token to mislead users. The practice also raises questions about whether UI designers will bake similar visual anchors into chat interfaces, reducing the need for manual token checks. Watching how AI providers handle custom instruction persistence and whether they expose native “drift alerts” will indicate whether such community hacks evolve into built‑in safety features.