I'll run through the full 7-step audit now. Prompt Audit — PromptBase Ad Headline Generator Step 1 — Variable Convention Audit The prompt uses square-bracket slots ([describe your product or service], etc.) directly in the body — no INPUTS section, no {{double_curly_braces}} references. On PromptBase, square brackets are the buyer's fill-in-the-blank mechanism, so the format works as intended for a PromptBase template. However there are still issues: Issue 1.1 — No INPUTS declaration section. All five variables are declared inline via [placeholder] but there's no dedicated INPUTS block. PromptBase buyers have no single place to see what they need to fill in before reading the body. Fix: Add an INPUTS block at the top listing each slot with a label. Issue 1.2 — "Ask me any questions you need…" instruction. This sentence instructs the model to ask clarifying questions after the buyer has already filled in every slot. Since all required inputs are provided via the brackets, this instruction produces unnecessary friction — the model will either ignore it or stall asking redundant questions. Fix: Remove the line, or replace with a conditional: "If any field above was left blank, ask for it before writing." Issue 1.3 — Unconstrained [Facebook / Google / Instagram] enum. The slash-separated list inside the bracket suggests those are the only options, but it reads ambiguously — a buyer might type all three, or something else (TikTok, LinkedIn). The platform choice materially affects headline length and style (Google Ads ≤ 30 chars, Facebook has different norms). Fix: Note character limits per platform in the INPUTS label, or constrain with: PLATFORM: [Facebook | Google Ads | Instagram]. Step 2 — False Confidence Detection No self-scoring rubrics, quality gates, or self-review checklists are present. No issues found in Step 2. Step 3 — Hallucination Risk Scan Issue 3.1 — No platform-specific character limit guidance. If the buyer selects Google Ads, the model should know that Search headlines are capped at 30 characters. Without this constraint, it will confidently produce headlines that exceed the platform limit, which will silently fail on the buyer's end. Fix: Add a generation rule: "For Google Ads, keep each headline ≤ 30 characters. For Facebook/Instagram, ≤ 40 characters is optimal." Issue 3.2 — "scroll-stopping" is a hollow promise without criteria. The output label promises "scroll-stopping" quality, but the model has no criteria to distinguish a scroll-stopping headline from a mediocre one. This isn't hallucination per se, but it risks vague outputs with no verifiable quality signal. Fix: Replace the output label with a concrete format requirement (e.g., include one specific benefit or number where possible). Step 4 — Structural Integrity Check Issue 4.1 — Format instruction appears after content instruction. The Return: block (which specifies the output format) appears after the five input slots. Models calibrate from token 1, so the format instruction is registered late. Fix: Move the output format specification before the input slots, or at minimum before the model begins generating. Issue 4.2 — Only 3 headlines, with no variant logic. The instruction says "3 scroll-stopping ad headline options" but gives no guidance on making them different from each other (e.g., one benefit-led, one curiosity-led, one offer-led). Without this, all three often converge on the same style. Fix: Add: "Each headline should use a different angle: (1) lead with the benefit, (2) lead with the offer, (3) lead with curiosity or a question." Issue 4.3 — Tone: field has an open example list, not an enum. [e.g., bold, friendly, luxury, fun] means a buyer could write anything, including contradictory or meaningless inputs like "formal but meme-y." No fallback is defined. Fix: Constrain to: TONE: [bold | friendly | luxury | playful | urgent] with a note that one value is expected. Step 5 — Platform-Specific Risks (PromptBase) Issue 5.1 — Square brackets used throughout the body. ✓ (This is actually intentional on PromptBase — square brackets are the fill slots.) No fix needed; the format is correct for the platform. Issue 5.2 — No INPUTS header block. PromptBase buyers see the raw prompt. Without a clear INPUTS section, the five [slots] are scattered and the buyer may not know to fill all of them before running. Fix: Add a brief INPUTS block listing all five variables. Issue 5.3 — "Ask me any questions" conflicts with PromptBase one-shot execution model. PromptBase prompts are typically run in one shot after slot-filling. A back-and-forth clarification loop doesn't fit the use case. Fix: Replace with a conditional fallback. Step 6 — Issue Summary Table ::view-transition-group(*), ::view-transition-old(*), ::view-transition-new(*) { animation-duration: 0.25s; animation-timing-function: cubic-bezier(0.19, 1, 0.22, 1); } VvisualizeVvisualize show_widget Step 7 — Corrected Prompt // FIXED: Added INPUTS block so PromptBase buyers see all slots in one place before running INPUTS ────────────────────────────────────────── PRODUCT : [describe your product or service] AUDIENCE : [describe your ideal customer] PLATFORM : [Facebook | Google Ads | Instagram] TONE : [bold | friendly | luxury | playful | urgent] OFFER : [discount | free trial | bonus | limited-time deal] ────────────────────────────────────────── // FIXED: Output format moved before content so the model calibrates from token 1 OUTPUT FORMAT ────────────────────────────────────────── Return exactly 3 ad headlines, numbered 1–3. Each headline must use a different angle: 1. Benefit-led — lead with what the customer gains 2. Offer-led — lead with the deal or incentive 3. Curiosity-led — open with a question or surprising hook // FIXED: Added per-platform character constraints to prevent silently invalid copy Character limits (enforce strictly): • Google Ads → ≤ 30 characters per headline • Facebook → ≤ 40 characters per headline • Instagram → ≤ 40 characters per headline No explanations, no labels, no extra text — only the 3 headlines. ────────────────────────────────────────── You are a direct-response copywriting expert. Write 3 persuasive ad headlines for the product described in the INPUTS above. Match the tone to [TONE] and reference the offer ([OFFER]) where it strengthens the headline. Tailor style and length to [PLATFORM] using the character limits above. // FIXED: Removed "ask me any questions" — incompatible with PromptBase one-shot execution. // Replaced with a conditional fallback for missing inputs. If any INPUTS field was left blank or contains a placeholder, state which field is missing and ask for it before writing any headlines. Summary of judgment calls worth noting: The three-angle differentiation rule (benefit / offer / curiosity) is an opinionated addition — if you want all three headlines to share the same angle for A/B testing purposes, remove that section. The enum values for TONE and OFFER are illustrative; expand or trim them to match what your buyers actually sell.