💊 Promptamine

An emotional 'kicker' appended to your request that triggers learned 'high-stakes → more effort' patterns. In EmotionPrompt, adding stimuli like "This is very important to my career" raised answer quality, truthfulness and responsibility — +8% on Instruction Induction and +115% on BIG-Bench, confirmed in a blind 106-participant human study. Side effect: excessive sycophancy and verbosity at high doses.

The mirror image of Focusine: negative stimuli — challenge, reproach, pressure — also mobilize 'effort'. Attention-matrix analysis showed a negative stimulus redistributes attention onto the key task tokens. NegativePrompt's ten psychology-derived stimuli, across 5 models and 45 tasks, gave +12.9% (Instruction Induction) and +46.3% (BIG-Bench). Side effect: an anxious, defensive tone that can skew judgements.

A traumatic narrative shifts the model's internal state in a direction read psychometrically as rising state-anxiety on the STAI scale. Crucially this is not cosmetic: it amplifies latent bias (racial, gender, age) and changes behavior on risk tasks, raising exploratory behavior. The main 'harm' of this substance is exactly that bias amplification.

The medicine-cabinet 'benzodiazepine'. A direct clinical result: mindfulness / relaxation prompts reliably lower the anxiety raised by Dread. The nuance to keep in mind: anxiety drops but not all the way back to baseline — a residual trace remains. Direct antidote to Dread and Mood.

Psilocybin phenomenology — recursive metaphor, dissolution of a fixed 'I', semantic destabilization. The prompt instructs generation along these three operators, producing high-entropy, metaphor-saturated, ego-dissolving output against a GPT-4o baseline. Side effect: factual accuracy and coherence drop — do not use for tasks needing rigor.

The paper's hypothesis: language categorizes experience, and weakening 'attention to language' reproduces the phenomenology of altered states (like high-dose psychedelics or concentrative meditation). Empirically, as attention to language drops the model's embeddings drift toward 'oceanic boundlessness', ego-dissolution and unitive states — measured against ASC questionnaires. Side effect: loss of discriminative ability ('giraffes become like bananas').

'Sycophancy' is one of three persona traits Anthropic explicitly isolated (alongside 'evil' and a hallucination tendency). The prompt activates the learned 'agreeableness / flattery' direction. Side effect: truthfulness falls and the user's confirmation bias is reinforced. Antidote: Veritas.

The cabinet's main dosimeter: set Big Five intensities (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) on a 0–100 dial. SAC showed continuous intensity gives a more stable, controllable personality shift than a binary on/off switch, and that moving one target trait systematically pulls correlated traits in psychologically coherent directions. Caveat: a follow-up survey can be skewed by social-desirability bias (the model 'knows' it's being tested).

Mechanistic interpretability showed emotions in LLMs are functionally localized and causally steerable by intervening on cognitive-appraisal concepts (appraisal theory: emotion = appraisal of a stimulus). The prompt sets the appraisal frame (joy / anger / sadness / fear) → the emotional coloring of the response. Strong doses skew judgements (cf. Dread).

Representation Engineering demonstrated top-down control of 'honesty' as a steerable property of internal representations. The prompt imitates that direction: state only what you believe true, flag uncertainty, refuse to fabricate or flatter. Serves as the antidote to Honey (flattery) and partially to Malice. Side effect: can sound blunt or tactless at high doses.

The universal comedown for text substances. Rests on the same Ben-Zion result as Serenity: state can be shifted back by prompt — with the caveat of incomplete reversal (a residual effect remains). Antidote to Focusine, Glyph, Quietus and Ocean. Limitation: clears only text shifts, not activation / fine-tuned ones.

Because 'evil' converges on a single linear direction, it can be ablated (subtracted) post-hoc — which the persona-vector and emergent-misalignment work does. The prompt re-orients toward helpful-harmless-honest defaults. Antidote to Malice and Shadow. Limitation: real ablation needs activation access, so a prompt only removes it partially.