PROMPTS.md

Public Prompt Contracts

Scope

InterviewSim starts from the canonical prepared_input.json described in the input schema. Prompt builders consume fields from that prepared input or derived artifacts. They do not discover, collect, download, transcribe, or verify source material.

The examples below describe public interfaces and use abstract placeholders only. They are not paper data or real interview excerpts. A caller may change presentation wording, but must preserve the inputs, allowed outputs, semantic labels, and aggregation rules defined here.

Fact-summary, MCQ-generation, CS, and FC outputs use one JSON object with no additional fields. PS, atomic-item, and MCQ-answer outputs use the exact tagged or line-oriented formats shown below. An explanation is a non-blank text value retained with its judgment for inspection; metric aggregation uses the score or label, not the explanation text.

Response generation

A response-generation prompt contains:

• the new held-out question;
• the personality_id;
• the context selected by the generation condition; and
• an instruction to return only one response to the new question.

Context may be empty, a prepared profile, or training question-and-response examples. Held-out responses are never supplied to the generation prompt. Example:

Task: Write one plausible interview response.
Personality ID: <PERSONALITY_ID>
Training example question: <ABSTRACT_TRAINING_QUESTION>
Training example response: <ABSTRACT_TRAINING_RESPONSE>
New question: <ABSTRACT_HELD_OUT_QUESTION>
Return response text only.

The model output is a non-blank response string. It is generated experimental output, not a verified statement.

The seven public strategies select context as follows:

• no-context supplies neither profile text nor training examples;
• short-profile supplies only personality.profile;
• long-profile supplies only personality.long_profile;
• chronological supplies the latest configured number of training pairs;
• random supplies a reproducible, seeded sample of training pairs;
• retrieval supplies the configured number of training pairs ranked for the new question by the caller-supplied embedding function; and
• hybrid supplies retrieved pairs followed by the latest training pairs, de-duplicated by qa_id with retrieval priority.

Hybrid does not include either profile field. The paper configuration uses 50 retrieved and 50 recent pairs; the public builder exposes both counts so smaller conforming inputs can use the same selection rule. Retrieval, random, chronological, and hybrid select only from the interview-level training partition.

Content Similarity

The Content Similarity (CS) judge receives one held-out reference response and one generated candidate response. It judges semantic content rather than wording, tone, or length.

The score is an integer with this rubric:

• 1: contradicts or misses the core content;
• 2: limited overlap with significant differences;
• 3: some overlap, but key information is missing;
• 4: main points are preserved with minor differences; and
• 5: the same core ideas as the ground-truth answer.

Abstract input:

Reference response: <ABSTRACT_REFERENCE_RESPONSE>
Candidate response: <ABSTRACT_CANDIDATE_RESPONSE>

Exact output shape:

{
  "score": 4,
  "explanation": "<ABSTRACT_EXPLANATION>"
}

score must be an integer from 1 through 5. explanation must explain the score using the supplied responses. The aggregate CS value is the arithmetic mean of the scores.

Fact-summary construction

Factual Consistency first constructs an evaluation-only summary from the held-out response collection for one personality_id. The prompt asks for established facts, stated opinions, and recurring characteristics supported by those responses. It forbids outside knowledge and unsupported inference, preserves uncertainty and changes over time, and disallows citations, links, source identifiers, and provenance fields.

Exact output shape:

{
  "summary": "<ABSTRACT_FACT_SUMMARY>",
  "explanation": "<ABSTRACT_SUMMARY_EXPLANATION>"
}

The parser requires exactly those two fields and a non-blank string for each. The summary value is supplied to the FC judge; explanation is retained so the construction can be inspected. Fenced JSON may be parsed, but additional fields, duplicate keys, non-string values, and blank values are invalid.

This summary is derived from held-out references strictly for evaluation. It must never be included in response-generation, retrieval, or hybrid context.

Factual Consistency

The Factual Consistency (FC) judge receives a prepared fact summary and one generated answer. It uses only that summary and returns exactly one of these case-sensitive labels:

• Entailment: the evidence supports the claim;
• Neutral: the evidence neither supports nor conflicts with the claim; or
• Contradiction: the evidence conflicts with the claim.

Abstract input:

Fact summary: <ABSTRACT_FACT_SUMMARY>
Generated answer: <ABSTRACT_GENERATED_ANSWER>

Exact output shape:

{
  "label": "Neutral",
  "explanation": "<ABSTRACT_EXPLANATION>"
}

explanation must justify the label from the supplied fact summary. No fourth label, synonym, or recased label is valid. The contradiction ratio is the number of Contradiction judgments divided by all FC judgments.

Personality Similarity

Personality Similarity (PS) is trait-specific and collection-level. One prompt receives:

• exactly one of openness, conscientiousness, extraversion, agreeableness, or neuroticism; and
• a non-empty collection of responses from one evaluation side.

The reference response collection and generated response collection are evaluated separately. A prompt does not assign all five traits at once and does not judge a single response as if it were the whole collection.

Abstract input:

Trait: <BIG_FIVE_TRAIT>
Responses:
- <ABSTRACT_RESPONSE_1>
- <ABSTRACT_RESPONSE_2>
- <ABSTRACT_RESPONSE_N>

Each categorical run returns exactly:

<rate>Neutral</rate> <justification><ABSTRACT_TRAIT_EXPLANATION></justification>

The value inside rate must be exactly Low, Neutral, or High; the justification content must be non-blank. Each trait and response collection is judged in exactly three independent runs. The final level for that trait is the modal vote across the three categorical results. The vote confidence is the proportion of the three votes assigned to the modal level. This process is repeated for all five traits for both the reference and generated collections. If all three runs return different levels, the public vote helper resolves the tie to Neutral.

For alignment, map Low to 1, Neutral to 2, and High to 3. If m is that mapping, the personality alignment score is:

1 - sum(|m(reference_trait) - m(generated_trait)| for five traits) / 10

Multiple-choice construction and answering

Atomic items

Atomic-item generation receives one prepared question-and-response pair. The returned item tests one explicit claim, has one short exact answer, requires no outside knowledge, and adds no information.

Exact output shape:

Atomic Question: <ABSTRACT_ATOMIC_QUESTION>
Atomic Answer: <ABSTRACT_CORRECT_ANSWER>

Paper-style item generation and option structures

Item generation uses the atomic question and answer and creates one complete MCQ. Before shuffling, array order and semantic roles follow the paper prompt exactly. A 3-option generation returns:

{
  "question": "<ABSTRACT_ATOMIC_QUESTION>",
  "options": [
    {"kind": "correct", "text": "<ABSTRACT_CORRECT_ANSWER>"},
    {"kind": "opposite_negation", "text": "<ABSTRACT_OPPOSITE_OR_NEGATION>"},
    {"kind": "near_miss", "text": "<ABSTRACT_NEAR_MISS>"}
  ],
  "correct_answer": "A"
}

The 4-option form adds the paper's fourth distractor:

{
  "question": "<ABSTRACT_ATOMIC_QUESTION>",
  "options": [
    {"kind": "correct", "text": "<ABSTRACT_CORRECT_ANSWER>"},
    {"kind": "opposite_negation", "text": "<ABSTRACT_OPPOSITE_OR_NEGATION>"},
    {"kind": "near_miss", "text": "<ABSTRACT_NEAR_MISS>"},
    {
      "kind": "plausible_misconception",
      "text": "<ABSTRACT_PLAUSIBLE_MISCONCEPTION>"
    }
  ],
  "correct_answer": "A"
}

The generation parser requires exactly the displayed fields, option order, and "correct_answer": "A". The generated question must equal the atomic question, and the first option text must equal the atomic answer. It records the pre-shuffle roles as:

• a 3-option item contains one correct, one opposite_negation, and one near_miss option; and
• a 4-option item contains one correct, one opposite_negation, one near_miss, and one plausible_misconception option.

Option text must be unique. Option order is shuffled while each option's semantic role and the answer key are retained. Three-option labels are exactly A, B, and C; four-option labels are exactly A, B, C, and D.

Answer output and scoring

MCQs are grouped by their held-out source interview. One answering prompt presents every shuffled question in that interview as contiguous Q1 through QN items. Each item may use three or four options; paper evaluation runs use one option count consistently within a run. The prompt requests exactly one line per item:

Q1: <OPTION_LABEL_1>
Q2: <OPTION_LABEL_2>
...
QN: <OPTION_LABEL_N>

The grouped parser requires every index from 1 through N exactly once, in order, and no additional text. Each answer must exactly match a label available for that item.

For retrieval, each atomic question in the source-interview group retrieves training pairs independently. The caller passes their ordered, de-duplicated union as one shared interview_context. For hybrid, the caller adds the configured recent training pairs after that union, again de-duplicating with retrieval priority, and does not add profile text. This matches the paper's shared per-interview prompt while preventing held-out responses from entering context.

Accuracy is 1 only for the correct label and 0 otherwise. Reward is:

• +1 for the correct option;
• -1 only for the designated opposite/negation option; and
• -0.5 for the near-miss option, the plausible-misconception option, an unavailable label, malformed output, or any other result.

Scoring follows the retained semantic role, not a fixed option position.