Skip to content

Built-in prompts

Taters ships with a small library of prompts for the "Generate responses with an LLM" step. Like the built-in dictionaries, you do not have to download or import them: the first time Taters needs your library it copies them into ~/.taters/library/prompts/, and from then on they appear wherever the step asks you which prompts to run, and under Settings and tools → Manage Taters data → Manage language model prompts.

They are ordinary .csv files (one role,content row per turn), so you can open one, see exactly what the model is told, copy it, change a definition, and import your own version alongside the original. Upgrading Taters keeps them current the same way it does the dictionaries: a release that adds or corrects a prompt brings it to your library, but a built-in you deleted stays deleted and a copy you edited is never overwritten.

These are not validated measures (yet)

The theory-grounded prompts below were written for Taters. Each one is built on a published theory and, where there is one, its standard measure, and the definitions follow those sources closely. None of them has been validated against self-report the way a few published prompts have (Peters & Matz, 2024; Wright et al., 2026; Hur et al., 2024). Before you treat a score as a measurement, check its convergent and discriminant validity on your own data, and report which prompt you used. The run's report quotes every prompt verbatim, which takes care of the second part.

Cite what you use

Every prompt here stands on somebody's theory, and usually on years of measurement work. If a score ends up in a paper, cite the sources listed with it alongside Taters itself.

Running them well

All of the theory-grounded prompts ask the model for a JSON object, so turn on llm_parse_json in the step's settings: each key in the reply becomes a column, named after the prompt (the Moral foundations (MFQ-2) prompt's care score lands in Moral_foundations_MFQ_2__care). Where a prompt asks for an object inside the object, like the Big Five's {"extraversion": {"evidence": ..., "score": ...}}, each inner value gets a column of its own: Big_Five_BFI_2__extraversion_score and Big_Five_BFI_2__extraversion_evidence. Lists (the agency and communion themes) are kept as JSON text in one cell.

The rest follows the practice in Rathje et al. (2024) and Wright et al. (2026), and most of it is what Taters does anyway:

  • Temperature 0, which is the default. Same model, same settings, same reply.
  • One text per conversation. Every text is asked about from scratch; the only earlier turns the model sees are the prompt's own worked examples.
  • Check the parse. A reply that does not parse is a row with blank columns and parse_ok 0, and the report says how many replies each key appeared in. Taters does not re-ask the model when a reply fails to parse or a score is out of range, so look at those rows before you analyze.
  • Scores about people need many texts. A prompt that rates the author (the Big Five, values) is most trustworthy averaged over many texts from the same person, and better still over repeated runs and several models: Wright et al. (2026) found the average across models agreed with self-report best of all. Taters' feature averaging handles the first part; running the same pipeline with a second model handles the third.

Theory-grounded prompts

Big Five (BFI-2)

Rates the author. 1–5 per domain, with a short note on the evidence for each. Columns: extraversion_score, agreeableness_score, conscientiousness_score, negative_emotionality_score and open_mindedness_score, each with its _evidence note.

The five-factor model, using the domain definitions of the BFI-2: Extraversion, Agreeableness, Conscientiousness, Negative Emotionality and Open-Mindedness. The model is told to rate typical, enduring tendencies rather than the mood of the moment, to give 3 when the text says nothing about a domain, and never to lean on assumptions about age, gender or group membership. This is a holistic rating of each domain. The best-validated approach so far is Wright et al.'s item-level method instead: one inventory item per call, answered as the author would answer it, then scored with the inventory's own key. Taters does not do that one (yet).

  • Soto, C. J., & John, O. P. (2017). The next Big Five Inventory (BFI-2): Developing and assessing a hierarchical model with 15 facets to enhance bandwidth, fidelity, and predictive power. Journal of Personality and Social Psychology, 113(1), 117–143. https://doi.org/10.1037/pspp0000096
  • Peters, H., & Matz, S. C. (2024). Large language models can infer psychological dispositions of social media users. PNAS Nexus, 3(6), pgae231. https://doi.org/10.1093/pnasnexus/pgae231
  • Wright, A. G. C., Ringwald, W. R., Vize, C. E., & Eichstaedt, J. C. (2026). Assessing personality using zero-shot generative AI scoring of brief open-ended text. Nature Human Behaviour, 10(3), 541–555. https://doi.org/10.1038/s41562-025-02389-x

Moral foundations (MFQ-2)

Rates each text. 0–3 per foundation, with three worked examples. Columns: care, equality, proportionality, loyalty, authority, purity, and a rationale.

Moral Foundations Theory as revised for the MFQ-2, which splits Fairness into Equality and Proportionality for six foundations in all. Following how moral language is annotated by hand (as in the Moral Foundations Reddit Corpus), a foundation counts only when the author makes a moral evaluation that rests on it: approval, condemnation, obligation or outrage. A topic that merely comes up scores 0, which is the line keyword-based moral dictionaries cannot draw.

  • Graham, J., Haidt, J., Koleva, S., Motyl, M., Iyer, R., Wojcik, S. P., & Ditto, P. H. (2013). Moral foundations theory: The pragmatic validity of moral pluralism. Advances in Experimental Social Psychology, 47, 55–130. https://doi.org/10.1016/B978-0-12-407236-7.00002-4
  • Atari, M., Haidt, J., Graham, J., Koleva, S., Stevens, S. T., & Dehghani, M. (2023). Morality beyond the WEIRD: How the nomological network of morality varies across cultures. Journal of Personality and Social Psychology, 125(5), 1157–1188. https://doi.org/10.1037/pspp0000470
  • Trager, J., et al. (2022). The Moral Foundations Reddit Corpus (arXiv:2208.05545). arXiv. https://doi.org/10.48550/arXiv.2208.05545

Schwartz basic values

Rates the author; works on single texts but means most averaged over a person's texts. 0–3 per value. Columns: self_direction, stimulation, hedonism, achievement, power, security, conformity, tradition, benevolence, universalism, and a rationale.

Schwartz's ten motivationally distinct values, arranged in a circle where neighbors are compatible and opposites conflict (self-direction against conformity, power against universalism). Values show in what a person pursues, praises, defends, or criticizes others for lacking, so the model is told to rate what the author values rather than what they describe, and not to give opposing values both high scores without clear evidence for each. If you follow standard Schwartz scoring, center each person's ten scores on their own mean (ipsatize) before comparing people.

  • Schwartz, S. H. (1992). Universals in the content and structure of values: Theoretical advances and empirical tests in 20 countries. Advances in Experimental Social Psychology, 25, 1–65. https://doi.org/10.1016/S0065-2601(08)60281-6
  • Schwartz, S. H., et al. (2012). Refining the theory of basic individual values. Journal of Personality and Social Psychology, 103(4), 663–688. https://doi.org/10.1037/a0029393

Agency and communion

Rates each text; written for personal narratives. 0–3 for each, plus the themes found, with two worked examples. Columns: agency, communion, agency_themes and communion_themes (JSON lists), and a rationale.

Bakan's duality of agency (self-assertion, mastery, separation) and communion (connection, intimacy, union), coded with the themes of McAdams et al.'s (1996) system for autobiographical scenes: four agentic themes (self-mastery, status/victory, achievement/responsibility, empowerment) and four communal ones (love/friendship, dialogue, caring/help, unity/togetherness). The two are independent, so a story can be high on both. The same two dimensions run through the dual perspective model of social cognition. If you want the word-counting version instead, the built-in Big Two dictionary measures the same pair.

  • Bakan, D. (1966). The duality of human existence: Isolation and communion in Western man. Rand McNally.
  • McAdams, D. P., Hoffman, B. J., Mansfield, E. D., & Day, R. (1996). Themes of agency and communion in significant autobiographical scenes. Journal of Personality, 64(2), 339–377. https://doi.org/10.1111/j.1467-6494.1996.tb00514.x
  • Abele, A. E., & Wojciszke, B. (2014). Communal and agentic content in social cognition: A dual perspective model. Advances in Experimental Social Psychology, 50, 195–255. https://doi.org/10.1016/B978-0-12-800284-1.00004-7

Depressive symptoms (PHQ-9)

Rates each text (or all of one person's texts joined into one). 0–3 per symptom, with whether the evidence was stated or inferred, and a short quote. Columns: for each of anhedonia, depressed_mood, sleep, fatigue, appetite, worthlessness_guilt, concentration, psychomotor and suicidal_ideation, a _score, an _evidence (explicit, inferred or none) and an _excerpt.

The nine PHQ-9 items, which map onto the DSM criteria for a major depressive episode, scored one by one on the PHQ-9's own frequency anchors. The model scores what the author states first, then infers the rest cautiously, and says which is which, following Ganesan et al.'s analysis of how GPT models read depression in text. The prompt tells the model not to add up a total: add the nine _score columns yourself for the 0–27 total, and use the usual 5 / 10 / 15 / 20 cutoffs for the severity bands.

This is a research scoring aid, never a diagnosis. Route any suicidal_ideation_score above 0 to a person under your IRB's safety protocol. And do not read a 0 there as evidence of absence: Ganesan et al. found GPT-4 reluctant to infer suicidality unless it was stated outright.

Core affect

Rates each text. Positive 0–10 and negative 0–10, rated separately, and arousal 1–9. Columns: positive, negative, arousal, and a rationale.

Russell's core affect, where feelings vary along valence and arousal, with one change that matters a great deal: valence is rated as two separate scales, positive and negative, following the bivariate model of evaluative space, so mixed feelings score high on both instead of averaging out to a "neutral" that is anything but. Arousal uses the anchors of the Self-Assessment Manikin. The model rates the feeling the author expresses, not whether the topic is good or bad, which is where generic sentiment prompts most often go wrong: a flat report of a tragedy is low on negative feeling, and a cheerful account of a hard day is high on positive. For a single valence score, subtract negative from positive, as Hur et al. (2024) did.

  • Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161–1178. https://doi.org/10.1037/h0077714
  • Russell, J. A. (2003). Core affect and the psychological construction of emotion. Psychological Review, 110(1), 145–172. https://doi.org/10.1037/0033-295X.110.1.145
  • Cacioppo, J. T., & Berntson, G. G. (1994). Relationship between attitudes and evaluative space: A critical review, with emphasis on the separability of positive and negative substrates. Psychological Bulletin, 115(3), 401–423. https://doi.org/10.1037/0033-2909.115.3.401
  • Bradley, M. M., & Lang, P. J. (1994). Measuring emotion: The Self-Assessment Manikin and the semantic differential. Journal of Behavior Therapy and Experimental Psychiatry, 25(1), 49–59. https://doi.org/10.1016/0005-7916(94)90063-9
  • Hur, J. K., Heffner, J., Feng, G. W., Joormann, J., & Rutledge, R. B. (2024). Language sentiment predicts changes in depressive symptoms. Proceedings of the National Academy of Sciences, 121(39), e2321321121. https://doi.org/10.1073/pnas.2321321121

Discrete emotions (appraisal)

Rates each text. 0–3 per emotion; more than one can be present. Three worked examples. Columns: anger, fear, sadness, disgust, joy, surprise, and a rationale.

Ekman's six basic emotions, but each one is defined by its core relational theme in Lazarus's sense: the way the author sees their situation that gives rise to it, rather than the words they use for it. That is what lets the model tell anger (a wrong, and someone to blame) from sadness (a loss that cannot be undone) when the vocabulary overlaps. Ratings are from the writer's perspective, not the reader's. If your research needs the self-conscious emotions (pride, shame, guilt), copy the prompt and add them with Lazarus's themes.

  • Ekman, P. (1992). An argument for basic emotions. Cognition and Emotion, 6(3–4), 169–200. https://doi.org/10.1080/02699939208411068
  • Lazarus, R. S. (1991). Emotion and adaptation. Oxford University Press.
  • Rathje, S., Mirea, D.-M., Sucholutsky, I., Marjieh, R., Robertson, C. E., & Van Bavel, J. J. (2024). GPT is an effective tool for multilingual psychological text analysis. Proceedings of the National Academy of Sciences, 121(34), e2308950121. https://doi.org/10.1073/pnas.2308950121

Simple starting points

Three plainer prompts ship as well. They are there to show the format and to get a first run going, not as measures.

Sentiment 1-7

Rates each text from 1 (very negative) to 7 (very positive), and asks for the number only. With llm_parse_json on, the bare number becomes the Sentiment_1_7__value column; off, it is the reply. For a measure of feeling, Core affect is the better choice.

Emotion with reason

The dominant emotion (joy, anger, sadness, fear, or none), its intensity from 1 to 5, and a one-sentence reason, as JSON, with two worked examples. Discrete emotions (appraisal) is the theory-grounded version.

One-sentence summary

A plain one-sentence summary of each text, in the writer's own terms. Free text, nothing to parse.