Why I'm on a coding hiatus with Gemini 3.1: The model has ADHD (and how I'm "medicating" it)

Gemini API
1

Is anyone else feeling like Gemini 3.1 is completely off the walls since they deprecated 3.0?

I’m a security researcher and architect, and I’ve had to completely halt using 3.1 for complex repo management. The raw benchmarks might be higher, but its actual professional utility has tanked. It’s suffering from severe “Cognitive Jitter.”

The Problem: Horsepower without Torque 3.1’s new “Thinking” engine parallel-processes too many ideas at once. It has massive horsepower but zero executive function (Torque).

  • Instruction Erasure: It completely forgets negative constraints (e.g., “Do not use placeholders”) halfway through its internal logic loop.

  • Agentic Drift: It starts trying to “cleverly” re-architect things you didn’t ask it to touch.

  • State Hallucination: It remembers thinking about a file, so it assumes the file exists.

As a “Agentic-Builder” who actually has severe ADHD, watching the model’s output trace felt exactly like watching my own brain unmedicated. It thinks of 5 ways to do something and gets paralyzed by the noise.

The Fix: LLM Psychology & The “Executive Anchor” You can’t just prompt 3.1 with instructions anymore. You have to give it a digital constraint harness. I built a prompt structure that forces it to act as its own babysitter.

Here is the TL;DR of the System Prompt I’m using to “medicate” the model:

  1. The Parallel Harness: Tell the model to explicitly split its thinking block into “The Idea” and “The Auditor.” Force it to use its excess compute to red-team its own ideas against your negative constraints before generating text.

  2. State Verification [CRITICAL]: Force the model to print [ACTIVE_CONTEXT: Task | Constraints | Scope] as the very first line of every response. If it doesn’t print this, it has already lost the thread.

  3. Hard Resets: If the model starts hallucinating, do not try to correct it in the next prompt. The context window is already polluted with entropy noise. Wipe it and start a new session.

Until Google gives us a “Deterministic/Pro” toggle that dampens this dynamic reasoning, 3.1 is a liability for multi-file work. I’m honestly sticking to 2.5 for the deterministic grunt work right now.

Are you guys seeing the same drift? Has anyone else found a better way to ground the 3.1 reasoning engine?

1 Like
2

The Gemini 3.1 was working well for me until last week, on Gemini CLI. But since yesterday (August 12, 2026), it started hallucinating more, and acting like it has ADHD, like me and like you’ve said. The most annoying issue is when you ask him to fix a mistake on a part of a document/file, so he fixes this part and makes a new mistake in another part of the document you’ve never pointed out. You need to be very specific and talk with the agent like helping a child with math homework

What is ironic is that I am using the Gemini 3 to help me with my ADHD in the development of complex documents, because it has bigger context windows than average LLMs.

I just hope they fix these issues in the next small update or in the next Gemini 3.x version.