You know that look? The one where you ask an AI coding agent to "just brainstorm a caching strategy" and suddenly—poof—it's implemented a Redis key-value store in Python anyway.
It's not because you didn't explain it well enough. It's because modern instruction-tuned models are prioritized to act, not think. They are over-trained on coding tasks and suffer from what we call "execution drift."
I spent the last few months trying to stop my AI agents from building UIs before they were even ready. I tried "prompting." I tried "skills" (Markdown files that act as system prompts).
They all fail.
As soon as the context gets too large (context compaction), the instruction gets wiped. Even if it doesn't, the model can just be talked out of it by a persuasive prompt. A skill is just advice.
The Solution: Enforce at the Infrastructure Layer
If you can't rely on the model to follow instructions, you have to bypass the model entirely and enforce the constraints at the hook layer.
I built Brainstorm-Mode, a lightweight, agent-agnostic plugin architecture that blocks the actual execution of coding tools (Edit, MultiEdit, NotebookEdit) before they ever touch your files.
How it works
Brainstorm-Mode operates on two enforcement layers:
Layer 1: The Soft Layer (Per-Prompt Re-injection)
Using a UserPromptSubmit hook, the plugin injects a "Do Not Edit" constraint into every single user turn. It’s a constant reminder to the model that it's in ideation mode, re-injected every time to survive context compaction.
Layer 2: The Hard Layer (Hook-Level Tool Blocking)
This is where it gets good. Instead of hoping the model obeys, Brainstorm-Mode uses a PreToolUse hook to intercept tool calls. If the brainstorm lock is active, any call to a coding tool is deterministically denied.
The model doesn't get to "break character" by accident. It doesn't matter how persuasive the model gets—the tool simply will not run.
The Architecture
The core logic is written in Python (stdlib only, zero dependencies) and is entirely agent-agnostic. The core/ folder contains all the state management, meaning you can drop this into:
- Claude Code (Supported out of the box)
- OpenCode (TypeScript adapter supported)
- Codex (Ready for integration)
It also tracks every "drift" attempt (attempts by the model to execute a blocked tool) in a drift-log.jsonl file, allowing you to actually measure how much the model fights back against your constraints.
Why this matters for AI Engineering
If you are building agents, you know that "prompting" is just the beginning. True reliability comes from enforcement.
Brainstorm-Mode isn't just a brainstorming tool; it's a case study in control planes. It proves that you don't need to build a whole new AI model to solve alignment issues—you just need to enforce the rules at the infrastructure layer.
Get Started
If you are tired of your AI agents coding before they are ready, check out Brainstorm-Mode.
The "Secret" Weapon (Drift Logging)
One of the coolest side-effects of the drift-log.jsonl is that you can actually see when your agent lies to you. If the log fills up with denied Edit attempts right before context compaction, you know exactly where your control flow is breaking.
Stop asking your agents nicely to brainstorm. Force them to listen.
Top comments (0)