The Prompt Library— Detecting Compliance Risks ── Ten working prompts that screen for exaggeration, off-label claims, and missing fair balance before human review

01Why "Is this a violation?" is the wrong question

The first instinct to abandon when using generative AI for compliance checks is letting the model rule on pass or fail. A language model picks the next likely word; it is not applying the legal elements of Article 66 of the Pharmaceutical Act. Ask "is this a violation?" and it returns a confident-sounding verdict drawn from the tone of the text. That verdict has no verifiable foundation.

The correct use is the AI as a net for candidates. The final ruling belongs to a human reviewer (regulatory affairs) who checks the primary sources. The model's job is limited to surfacing every passage worth a look and attaching a hypothesis for why it is suspect. Accuracy improves when you make it name which provision of the four-tier regulatory map from Part 3 (the Act, the Advertising Standards, the MSA guidelines, the JPMA Code) each candidate offends.

02The base block — fixing role, norms, and output format

Before any specific detection prompt, set up one "base block" that sits at the top of every run. Fixing the role, the referenced norms, and the output schema, then having later prompts inherit that block, reduces output drift and hallucination.

The base block has four parts: (1) state the AI's role as a candidate detector, not a final judge; (2) list the norms in scope (Act §66/§68/§68-2, the Advertising Standards, the MSA guidelines, the JPMA Code); (3) fix the output to a fixed table; (4) write explicit bans — no findings without textual basis, no inventing article numbers.

03Prompts 1–3: detecting exaggeration (Act §66)

Exaggerated advertising is barred by Article 66 — overstated claims about efficacy or safety, whether explicit or implied. Catch superlatives, absolutes, implied head-to-head comparison, and over-reading of data.

Prompt 1, superlatives and absolutes: "Given the base block, extract every superlative (most / No.1 / only / unmatched), every absolute (always / certainly / cures), and every implication of total safety (no side effects / worry-free) made about efficacy or safety. Flag any suggestion that exceeds the approved indication."

Prompt 2, over-reading data: "For each figure and number, check whether the claim exceeds what the cited data show. Flag relative-risk reduction shown without the absolute, a secondary endpoint treated as the primary effect, or a subgroup result presented as the whole picture."

Prompt 3, implied exaggeration: "Where there is no direct claim but a photo, color, arrow, testimonial, or tagline overstates effect or safety, flag it — including the alt text and captions of visual elements."

04Prompts 4–5: unapproved and off-label (Act §68)

Advertising an unapproved product, and any suggestion beyond the approved indication (off-label), fall under Article 68. Article 66 (exaggeration) and Article 68 (unapproved) are distinct, so the prompts must instruct the model not to confuse the two types.

Prompt 4, matching the indication: "List every efficacy, indication, dosage, and target population the material claims. Then classify each as within the approved label, outside it (off-label), or undeterminable. I will supply the approved label separately. Do not fill in the label from memory or guesswork; if none is provided, write 'needs label cross-check'."

Prompt 5, development-stage or unapproved hints: "Detect any wording that suggests efficacy or usefulness for an unapproved product, indication, or formulation — efficacy claims dressed up as a pipeline introduction, congress-presentation data repurposed for promotion. Label these as §68 candidates (pre-approval advertising), kept separate from §66 (exaggeration)."

05Prompts 6–7: fair balance and citations (MSA / Advertising Standards)

Whether safety information (adverse reactions, contraindications, warnings) balances the efficacy claims — fair balance — sits at the heart of the MSA guidelines and the Advertising Standards. Alongside it, check that each claim has a stated, verifiable source.

Prompt 6, fair-balance check: "List efficacy statements and safety statements separately, then contrast their volume, visual prominence, and placement. Where efficacy is up front and large while safety is a small footnote, flag a fair-balance gap. Also flag any missing serious adverse reaction, contraindication, or warning."

Prompt 7, source and citation integrity: "For every claim, number, and figure, check whether a source is stated and whether source and claim agree (the material does not say what the source does not). Flag quantitative claims with no source, secondary citation, use of unpublished data, and undisclosed conflicts of interest. Do not fill in source content by guesswork."

06Prompts 8–10: cross-cutting and diff review

After the per-type prompts, run three cross-cutting ones: whole-piece consistency, the revision diff, and a self-check.

Prompt 8, audience fit: "Determine the intended reader (healthcare professional vs. general public) and check that the language level fits. Flag promotion of unapproved or prescription-only efficacy to a general audience, and over-simplified absolutes aimed at professionals."

Prompt 9, regulatory impact of a revision: "Given the old and new versions, extract only the changes and judge whether each introduces new regulatory risk (creeping exaggeration, broadened indication, reduced safety text). Flag every deletion or shrinkage of safety information."

Prompt 10, the AI's self-check: "Re-examine your own previous output. Retract, with reasons, (1) any finding with no textual basis, (2) any invented article number or statistic, and (3) any item given inflated confidence. Output only the surviving, confirmed list."

Prompt 10 cannot fully erase the weakness of generating and evaluating in one context (self-approval). That is exactly why the final gate is human. The AI's self-check is only an aid.

07Hallucination controls — design that distrusts the output

Language models will calmly fabricate plausible article numbers, statistics, and citations. Build that assumption in, with mechanisms that structurally doubt the output.

1. Force verbatim quotes: every finding must quote the material verbatim. A finding that cannot quote the text is likely a hallucination.
2. Verify article numbers: trust no § the AI cites until a human checks it against the text of the article. Confusing §66 / §68 / §68-2 is especially common.
3. Cut off outside knowledge: a human pastes the label, the approval data, and source PDFs into the prompt; never rely on the model's memory. Enforce "if not provided, write needs-human-check".
4. Require confidence: demand high/med/low on every item; review all low and medium, and spot-check the highs.
5. Lower the temperature: creativity is unwanted, so keep the generation parameters toward the deterministic end.

08Turning it into an SOP — workflow and review record

A prompt library becomes an SOP only once the steps for using it and the record format are settled. Here is how to fold the pre-screen into the QC stage at Gate C. Place it at the head of the QC stage among the ten phases defined in Part 2, lifecycle design (coming soon).

The record exists for audit and traceability. Fold the AI-screening log into the "who, when, on what basis" chain covered in Part 10, traceability (coming soon). Running something through AI is not an excuse for liability. What the record shows is only the fact that the AI raised candidates and a human ruled against the primary sources.