Tutorial-and-demonstration post applying Zou et al. 2023's representation-engineering methodology to Mistral-7B-Instruct-v0.1. Single-component PCA on contrastive prompt-pair hidden states across 13 layers (~1 minute of compute) yields per-layer control vectors that steer behavior at inference time without prompting or fine-tuning. Six worked demonstrations — psychedelic effects, laziness/diligence, political orientation, creativity, temporal perspective, and self-awareness — plus a comparison with prompt engineering showing that control-vector strength can be tuned independent of prompt wording (jailbreak/anti-jailbreak section). No quantitative evaluation; qualitative generated-text samples only — below the wiki's empirical-substantiation bar for findings. The post's significance is downstream: the accompanying repeng PyPI library is the most widely-used community implementation of contrastive-PCA control-vector extraction, and the post itself was a major vector by which the Zou et al. methodology propagated into community practice. One non-trivial empirical observation worth flagging: when the honesty vector is applied while the model evaluates a third party's behavior, it shifts the model's judgment of that person's honesty rather than the honesty of the model's own output — a within-pass entanglement of self-state and other-state representation that the post does not analyze further. The "Future Work" section anticipates applying SAE monosemantic features as cleaner inputs than superimposed activations, foreshadowing the SAE-based extensions later filed in the wiki (OpenAI SAE on GPT-4o EM, Persona Vectors). Theia Vogel: independent blogger, no institutional affiliation; vgel.me hosts both technical writing and fiction.