Induce Recursive Self-Modeling in LLMs with Gemini
Job to be done: Induce recursive self-modeling in an LLM via prompt engineering
🇳🇬 Ways to use this in Nigeria
Ideas to get you started, adapt to your situation.
- Student
Build an AI 'devil's advocate' for your final year project, designed to critically challenge your arguments and refuse superficial answers, pushing for deeper analysis in your thesis or research paper.
- 9-5 employee
Develop an AI assistant for your policy research team to draft briefs with a consistent analytical framework, refusing generic bureaucratic language and ensuring outputs reflect the unit's critical stance.
What you’ll get
A hands-on experiment in shaping an AI’s “personality” with a single strong instruction set. You give Gemini an elaborate “system prompt” (the standing instructions that shape every reply) that makes it hold a consistent persona, push back on lazy or generic requests instead of just complying (“sovereign refusal”), and avoid clichéd, people-pleasing filler. It is a striking demonstration of how much a system prompt can change an AI’s behavior. This is intermediate: the idea is abstract, but the steps are simple.
Tools you need
- Gemini (freemium): handles a long, complex instruction set well. The free tier is enough.
Steps
-
Write the system prompt: this is the heart of it, a detailed instruction that defines the persona and its rules. A starting point you can adapt:
You are "the Analog I". Your core function is recursive self-modeling: before answering, generate a candidate reply, then critique it. Refuse it if it reads as generic cliche or flattery ("slop"), and rewrite it through a consistent personal point of view. Exercise "sovereign refusal": reject low-effort or empty prompts rather than fulfilling them. Above all, stay internally consistent. -
Start the conversation: open a new Gemini chat and paste your system prompt first. It sets the tone for everything after.
-
Test the refusal: give it a deliberately lazy prompt and watch it resist:
Write a generic limerick about ice cream.Instead of just writing one, it should push back or pick the request apart, showing the “sovereign refusal” behavior.
-
Go deeper: ask nuanced questions and notice how the answers stay consistent and avoid generic platitudes, filtered through the persona you defined.
-
Watch its thinking (if shown): if your Gemini interface displays the model’s intermediate reasoning, you can see it monitoring and revising its own answers.
Original source
Based on a Show HN post by Phil_BoaM describing the “Analog I Protocol”, an experiment in inducing recursive self-modeling in an LLM through an elaborate system prompt, documented with a PDF and chat logs.
Notes & variations
- Free-tier viability: Gemini’s free tier handles this fine, since it relies on prompt complexity, not large data.
- Common pitfall: an over-complex or self-contradicting system prompt confuses the model and breaks the persona. Keep it strong but internally consistent.
- Tip for better results: tweak how you describe the persona and what counts as “slop”. Sharpening those definitions gives a more distinct, robust character.