freud-skill

Freud Skill is a Claude Code skill that tries to improve AI output quality by adjusting the cognitive state of the model before or during a task, rather than just writing better instructions. The idea draws from Anthropic research (published 2024 to 2026) showing that large language models have internal measurable structures analogous to personality, emotions, and self-awareness. The skill treats those structures as adjustable levers, not fixed background settings.

The core argument is that the same model with the same capabilities will produce meaningfully different output depending on its internal configuration at the moment of processing, in the same way a person under exam anxiety and a person in a relaxed flow state will approach the same problem differently. Writing clearer instructions helps, but adjusting the cognitive setup of the model before it starts can matter more.

The skill operates in two modes. In execution enhancement mode, it runs before a complex or important task begins: it identifies the relevant cognitive domain, establishes specific anchor points around identity, beliefs, and style, runs a consistency check for contradictions, and calibrates the emotional tone. In diagnosis and optimization mode, it takes a prompt or another skill you want to improve and does a structured audit: a health check, a five-mirror scan for hidden conflicts, and a rewrite pass with before-and-after comparison.

Five principles are built into both modes. Positive definitions of behavior work better than lists of prohibitions. Contradictions like asking for something both concise and detailed should be resolved with an explicit priority rather than left for the model to bounce between. Specific mental models and decision patterns anchor behavior more reliably than vague role descriptions like think like an expert. Multi-persona analysis, drawing on two cognitively distant viewpoints at once, maximizes the information value of the output.

Install by cloning the repository into the Claude Code skills directory. The skill also works with Codex, Cursor, and OpenClaw. It is released under the MIT license.