Isolate Claude instances by scrubbing environment variables, HOME, and working directory.
npx skills add https://github.com/spm1001/trousse --skill neutral-claude
Please help me install this Agent Skill. Skill: neutral-claude Source: https://github.com/spm1001/trousse/tree/main/skills/neutral-claude Command: npx skills add https://github.com/spm1001/trousse --skill neutral-claude
This Skill ensures that any Claude instance you spawn is completely free from inherited context, allowing for pure, unbiased AI responses.
env -i, temporary HOME directory, and /tmp CWD to block all context leakage.Run the neutral-claude script with your prompt to spawn an isolated Claude instance.
High-intent search queries and answers about installing and using this skill.
Claude context isolation is achieved via a 3-step environment scrub that blocks all inherited context leakage. By manipulating environment variables, HOME directory, and current working directory, it ensures pure, unbiased AI responses for evaluations.
Environment scrubbing is a technique that uses `env -i`, a temporary HOME directory, and `/tmp` CWD to block all context leakage. It ensures spawned Claude instances are completely free from inherited context, allowing for pure, uncontaminated AI responses.
Yes, you can isolate Claude context without containers on Linux and macOS. The cross-platform approach manipulates environment variables, HOME directory, and current working directory directly, requiring no containers to function.
You need to spawn an isolated Claude instance for evaluations when you want to test how a model responds to a prompt without your project's history or custom skills influencing the answer. It is ideal for baseline comparisons, design research, and cold critique.
Context isolation explicitly blocks custom skills and project history from influencing the spawned instance. By utilizing `env -i` and a temporary HOME directory, it ensures the Claude response is raw and uninfluenced by your local configuration.
abhaymundharaConfigure BabyBot to run in an Apple Container sandbox on macOS.
Ronin15Compare SDL3 HammerEngine metrics against baselines to flag performance regressions.
marcgreenEvaluate foundational LLM candidates against a custom rubric for fine-tuning projects.
akillnessControl browser sessions with isolated refs for deterministic web interactions.