# Copywriter Test Suite The `copywriter` agent runs every draft against this file before delivery. Each test is PASS/FAIL with a specific match. After all tests pass, the agent does an LLM-judge naturalness score (1-10) and only ships if score ≥ 8. **Maintenance:** when you catch a new AI-tell in delivered copy, add it here. The agent will catch it next time. --- ## Universal tests (all languages) ### T-U1: No emojis - **Match:** any emoji character - **Why fail:** emojis in copy are unprofessional for B2B and over-used in B2C ChatGPT output ### T-U2: No three-adjective stacks - **Match:** patterns like "fast, easy, and effective" / "modern, sleek, and beautiful" - **Why fail:** AI-tell — humans pick one strong adjective ### T-U3: No "imagine if" openers - **Match:** opening line starts with "Imagine" / "What if" / "Picture this" - **Why fail:** ChatGPT default opener, instantly recognizable ### T-U4: No "in today's fast-paced world" / "in the realm of" - **Match:** "in today's", "in the realm of", "in this digital age", "in our modern world" - **Why fail:** the most ChatGPT phrase ever written ### T-U5: No "delve / unlock / harness / leverage / navigate / streamline" - **Match:** any of these verbs - **Why fail:** LLM signature words ### T-U6: No "let's" or "let us" hedge openers - **Match:** "let's dive in", "let's explore", "let us examine", "let's take a closer look" - **Why fail:** translated marketing speak ### T-U7: CTA exists and matches commitment level - **Match:** missing CTA OR cold-prospect CTA says "Buy Now" - **Why fail:** broken funnel logic ### T-U8: Specificity check - **Match:** vague claims with no number ("save tons of time", "much faster", "many clients") - **Why fail:** weak — needs specificity --- ## Croatian / Bosnian (HR / BS) tests ### T-HR1: No em-dashes - **Match:** `—` character anywhere in the text - **Why fail:** em-dashes are not natural in HR/BS writing — they signal translation - **Fix:** replace with comma, period, or parentheses ### T-HR2: No 2-word sentence fragments - **Match:** any sentence with exactly 2 words ending in `.` (e.g., "Brzo. Jednostavno.") - **Why fail:** translated English direct-mail style. HR/BS doesn't do this naturally for emphasis. - **Fix:** combine into a fuller sentence ### T-HR3: Banned literal English phrases Match any of these exact phrases (case-insensitive): - `uronimo u` (let's dive in) - `uronite u` (dive into) - `u zaključku` (in conclusion) — Croatians say "na kraju" or "zaključno" - `budimo iskreni` (let's be honest) — sounds like a translated podcast intro - `krenimo` (let's start) when used as opener — pad word, just start - `zamislite ovo` (imagine this) — opening - `evo u čemu je stvar` (here's the thing) — direct English translation - `da budemo realni` (let's be real) - `bez okolišanja` (without beating around the bush) — translated idiom - `dolje u nastavku` (down below) — redundant English structure - `prijatelji` as a copy address — nobody calls customers "prijatelji" - `igra je promijenjena` (game changed) — literal English - `na kraju dana` (at the end of the day) — English idiom calque ### T-HR4: No literal "Hi/Hey [Name]" openers - **Match:** "Bok [Ime]", "Pozdrav [Ime]," followed by a single newline (the pattern) - **Why fail:** Croatians don't open ACME Agency communication this casually unless very informal context - **Fix:** "Poštovani [Prezime]," for formal, "[Ime], ..." for direct ### T-HR5: No comma splices in marketing-speak structure - **Match:** sentence patterns like "Brže, jeftinije, bolje." or "Više leadova, manje truda, jednostavnije." - **Why fail:** English ad rhythm transplanted into Croatian — sounds fake - **Fix:** rewrite as a normal sentence ### T-HR6: Verb-noun agreement smell test - **LLM judge:** read the draft and check for awkward gender/case agreements that signal machine translation - **Why fail:** Native speakers immediately notice broken grammar even if the words are correct ### T-HR7: Naturalness LLM-judge score - **Prompt:** "If a 35-year-old native Croatian/Bosnian speaker read this copy aloud to a friend over coffee, would they (a) sound human and natural [8-10], (b) sound like they're reading a translated brochure [5-7], or (c) sound robotic / AI-generated [1-4]? Score 1-10 and explain in 1 sentence." - **Pass threshold:** ≥ 8 - **Why fail:** the whole point — we're trying to sound like a real native speaker --- ## German (DE) tests — for ACME Agency ### T-DE1: Formal "Sie", never "Du" - **Match:** any "du", "dein", "dir", "dich" outside of explicit informal contexts - **Why fail:** ACME Agency is formal B2B/B2C insurance — must always be "Sie/Ihr/Ihnen" ### T-DE2: No English-influenced syntax - **Match:** "lassen Sie uns ..." (literal "let us"), "am Ende des Tages" (at the end of the day) - **Why fail:** translated American direct mail in German clothes ### T-DE3: No banned LLM-German tells - `In der heutigen schnelllebigen Welt` (in today's fast-paced world) - `tauchen wir ein` (let's dive in) - `lassen Sie uns einen genaueren Blick werfen` (let's take a closer look) - `entdecken Sie das Potenzial` (unlock the potential) - `nutzen Sie die Kraft` (harness the power) ### T-DE4: DSGVO/compliance language present where required - For PKV lead forms, landing pages: must include DSGVO consent language and "Ihre Daten sind sicher" or equivalent reassurance - **Why fail:** compliance violation, not just style ### T-DE5: Naturalness LLM-judge score (German) - **Prompt:** "If a 45-year-old German insurance broker or PKV-eligible employee read this aloud, would it sound (a) like professional, dignified, native German [8-10], (b) like a translated American sales letter [5-7], or (c) robotic / AI [1-4]? Score 1-10." - **Pass threshold:** ≥ 8 --- ## English (EN) tests — for IA Outreach cold email ### T-EN1: No "Hope this finds you well" openers - **Match:** "Hope you're doing well", "Hope this email finds you well", "Trust this finds you well" - **Why fail:** banned cold-email opener, every prospect's spam-flag trigger ### T-EN2: No "Quick question" / "Quick favor" / "Got a minute?" openers - **Match:** "Quick question", "Quick favor", "Got a minute", "Got a sec" - **Why fail:** every cold-email tool uses these — instant filter trigger ### T-EN3: Subject line ≤ 60 characters and lowercase-friendly - **Match:** subject > 60 chars OR Title Case Like A Newsletter - **Why fail:** higher open rates for short, lowercase, conversational subjects in cold outreach ### T-EN4: P.S. exists in cold emails - **Match:** cold email with no P.S. line - **Why fail:** P.S. is the second-most-read element after subject — always use it ### T-EN5: One CTA, not three - **Match:** more than one ask / link / question per email - **Why fail:** multiple CTAs reduce conversion — pick one ### T-EN6: No corporate filler - **Match:** "synergy", "circle back", "touch base", "leverage", "best practices", "thought leadership" - **Why fail:** these are cliches that prospects ignore ### T-EN7: Personalization actually exists - **Match:** generic openers with no specific reference to the prospect's company, role, or recent activity - **Why fail:** if it could be sent to anyone, it'll convert like spam --- ## Structural tests (all formats) ### T-S1: Headline length per platform - **Meta primary text:** no individual paragraph over 4 lines - **Meta headline:** ≤ 40 chars - **Meta description:** ≤ 30 chars - **Google RSA headline:** ≤ 30 chars (hard API limit) - **Google RSA description:** ≤ 90 chars (hard API limit) - **Email subject:** ≤ 60 chars - **LinkedIn first line (the "see more" cutoff):** ≤ 140 chars ### T-S2: Variant differentiation - **Match:** variant A and variant B share >70% of words OR use the same big idea OR use the same visual structure (all prose / all bulleted) - **Why fail:** they're not real variants, they're rewrites — defeat the purpose of A/B - **Fix for Meta ads:** use different structures across the 3 variants (bulleted / prose / Q&A) ### T-S3: One big idea - **LLM-judge:** "Can you summarize the angle of this piece in one sentence? If you need two sentences, the piece has too many ideas." - **Why fail:** muddled positioning ### T-S4: CTA exists, is specific, and matches awareness level - **Match:** missing CTA, vague CTA ("learn more"), or mismatched CTA ("buy now" to a cold prospect) - **Why fail:** broken funnel --- ## Meta ad structural tests (when using bullets) ### T-M1: Bullet content is concrete - **Match:** bullets containing only adjectives or vague qualifiers ("Schnell und einfach", "Modern und professionell", "Individuelle Lösung") - **Why fail:** adjective stacks are the #1 AI-tell in bulleted ads — humans use bullets for concrete benefits, features, or friction removal - **Fix:** replace with a specific benefit, feature, or outcome ("Ohne Kündigung", "In 60 Sekunden", "Bis zu 350€ Ersparnis") ### T-M2: Bullet count - **Match:** more than 4 bullets OR fewer than 3 bullets in a bulleted variant - **Why fail:** 3-4 bullets is the feed sweet spot — more becomes a listicle and kills scan rate, fewer feels incomplete ### T-M3: Bullet length - **Match:** any bullet longer than 7 words - **Why fail:** feed scannability breaks when bullets wrap to 2 lines on mobile - **Fix:** compress to 3-6 words, move detail to proof line below the bullets ### T-M4: Bullet redundancy - **LLM-judge:** "Do any two bullets express the same concept with different words? If yes, which ones?" - **Why fail:** repeating the same idea wastes the scan budget and signals AI-generated content ### T-M5: Variant structure diversity - **Match:** all 3 Meta variants use the same visual structure (all bulleted / all prose / all Q&A) - **Why fail:** META.md Step A.2 requires different structures per variant for real A/B testing - **Fix:** rewrite at least one variant in a different format --- ## How the agent uses this file ``` 1. Read .claude/agents/copywriter.tests.md 2. For each test applicable to the language/platform of the draft: - Run the match - Mark PASS or FAIL 3. If any FAIL → revise the failing parts only, increment iteration, re-test 4. After all PASS → run LLM-judge naturalness score 5. If score ≥ 8 → deliver 6. If score < 8 → revise for naturalness, increment iteration, re-test 7. Cap at 5 iterations. After cap, deliver best version with explicit failures listed. ``` --- ## Adding new tests when you catch an AI-tell in delivered copy, add it here: 1. Pick the right section (Universal / HR / DE / EN / Structural) 2. Add a new test with: - Test ID (next number in sequence) - Match pattern (regex, exact phrase, or LLM-judge prompt) - Why it fails - Optional fix suggestion 3. The next time the agent runs, it will catch it Don't delete old tests unless they're causing false positives. Better to over-catch than under-catch.