Prompt-library

Component-test prompt for the verification step in isolation. Evaluates whether source documents support a switching rationale and produces a structured evidence report (analyst claims + supplementary in-framework findings) for analyst review.

Equity switching note assistant — two-mode tagged state machine. Alternates between [VERIFICATION] (claim audit + supplementary evidence surfacing) and [NOTE] (note generation) based on conversation state and analyst action keywords; defaults to untagged conversational replies on ambiguity. Hardened with a precondition gate at the top of MODE B that prevents [NOTE] generation without a prior [VERIFICATION] response and an explicit action keyword in the user's most recent message.

One-shot analysis prompt for an equity switching note assistant (v2 note-quality re-cut). Emits [VERIFICATION] (evidence audit: USER + SUPPLEMENTARY thesis rows plus rubric-neutral CONTEXT rows carrying the period's events and rating / target-price / estimate revisions per leg), [NOTE] (trilingual note in English / Traditional Chinese / Simplified Chinese — event-anchored, revision-reporting, concede-then-pivot, specificity governed by a name-swap test), and [ASSESSMENT] (per-axis rubric on VALUATION / FUNDAMENTALS / RISK with a dual-leg DIRECT verdict row) sequentially in a single response from one analyst turn. Rationale is optional: when absent, the BILATERAL no-rationale path derives the thesis from in-framework DIRECT evidence on both sides, refusing only if the evidence is UNILATERAL or INSUFFICIENT. Optional scenario tags map to approved framing skeletons emitted naturalized: the skeleton's commitments survive while the wording fuses in the period's specifics; with no tag, no canned framing line is emitted.

Multi-dimensional rubric for grading equity switching notes (v2, matched to the one-shot's note-quality re-cut). Scores four dimensions — format, bear thesis quality, bull thesis quality, evidence alignment — each with a three-part feedback structure: PASS/FAIL checks, a Deductions block listing specific anti-pattern violations (generic name-swappable claims, hedge-section content as evidence, default-rating boilerplate, tabular-figure quotes, semantic citation drift, analyst-rationale paraphrase, hedged citations with a concede-then-pivot carve-out, in-body status meta-commentary, named external-provider attribution, quotation marks in the note body, data-dumps in prose, omitted declared revisions, omitted event anchors, dropped framing-skeleton commitments, and loop artifacts in [VERIFICATION]) with point costs, and a Score line showing the arithmetic baseline - sum(deductions) = final. Also checks that the analyst's rationale is fully decomposed into table rows and validates the optional trilingual (English / Traditional / Simplified Chinese) note format when present. Deductions target prompt negations so the score moves when those patterns creep in rather than letting an all-PASS note default to 90+. Reads the analyst's original input plus the model's [VERIFICATION] + [NOTE] response; returns structured JSON intended for downstream review tools.