The Ad-Visor 3000 — by Conductor AI Labs.
// find out if your creative is original without having to search for hours on end
Drop in a headline, body copy, or a script. Ad-Visor 3000 cross-references it against the actual advertising corpus, runs a four-dimensional originality analysis with cited evidence, and plays you the reactions — not a verdict, reactions — voiced by ElevenLabs. So you can hear what “derivative” actually sounds like.
The headline is “reimagine possible.” The visual is a person looking thoughtfully out a window. The strategy is “disrupt the category.” Everyone nods. The invoice clears. The work runs. Nothing happens.
Ad-Visor 3000 exists because you shouldn’t need to pay $200,000 to find out you just bought the eighth version of an idea your own category has already produced. The advertising industry has a discoverability problem — nobody has time to audit a deck against Cannes Lions winners, Meta Ad Library, TikTok Creative Center, D&AD, and the One Show. So derivative work gets through.
Ad-Visor 3000 does that audit in sixty seconds.
Submit an ad. Watch the verdict land with a 0–100 score, four dimension breakdowns, cited prior art with clickable source URLs, and three one-sentence reactions voiced by ElevenLabs.
Overall score = concept × 0.40 + language × 0.25 + strategy × 0.20 + execution × 0.15. Cross-dimension inheritance rules enforce downgrades — a verbatim language match drops every dimension, because a copy is a copy.
Is the core idea, metaphor, or narrative unique? The hardest and heaviest dimension. A winning execution on a tired concept is still derivative.
Are the exact words, phrases, and linguistic patterns original? Verbatim borrows kill the score — and drag every other dimension down with them.
Is the positioning or angle unique within this industry? “Disrupt the category” is fresh in beer and stale in fintech. Industry context is everything.
Is the format or approach novel for this media type? A film trick that’s been done a hundred times in film is derivative even if it’s never appeared in out-of-home.
After the score, Ad-Visor 3000 plays three one-sentence lines voiced by ElevenLabs (Adam, deep and authoritative, eleven_turbo_v2_5). Not judgments — what real people would say after seeing your ad. Tone is driven by the score.
Wow, that was really original — I hadn’t thought about it that way.
I’m actually sending this to someone.
You don’t see work like that very often.
Yeah, I liked it — felt fresh.
I’d probably remember that one.
It’s got something.
It’s fine, but I feel like I’ve seen it before.
Didn’t really stop me scrolling.
Close, but not quite there.
Every brand does this.
I tuned out pretty fast.
It just kind of washed over me.
Oh no, not this again.
I’ve seen this exact ad a hundred times.
Nobody’s going to talk about this.
A single-vector search on the headline is a toy. Ad-Visor 3000 decomposes every ad into four independent semantic facets, embeds each one, and searches across all four in parallel — plus a web sweep — before Claude ever renders a verdict.
Claude Sonnet 4.6 receives the submitted ad and breaks it into a core_concept, emotional_tone, structural_pattern, five semantic_variants (including the contrarian version, the competitor’s version, and the cliché version), four cross_language_queries (ES / FR / DE / universal), and five firecrawl_queries optimized for Google — with exact-phrase quoting and award show names (Cannes Lions, D&AD, One Show) baked in.
Four independent embeddings computed via Cohere embed-multilingual-v3.0: headline, concept, tone, structure. Multilingual so the search works across non-English corpora without a separate translation pass.
TurboPuffer — vector search on each of the four facets against the adjudge-ads namespace. Neon Postgres — text search for exact language matches. Firecrawl — multi-query web scrape using the five Google-optimized queries, then content extraction to pull ads out of scraped pages. Web-discovered ads are cached back to the corpus so the next query is faster.
Candidates sorted into four match buckets by type — concept_matches, language_matches, strategy_matches, execution_matches. Strategy buckets are industry-filtered; execution buckets are media-type filtered. Each bucket is dimension-specific evidence for the judgment pass.
Claude Sonnet 4.6 scores each dimension independently against its evidence bucket, applies cross-dimension inheritance rules (verbatim language match → every dimension down), and returns a JSON verdict with 1–3 fully-cited evidence items per dimension — brand, agency, year, country, language, similarity percentage, specific overlap description, and source URL.
Score threshold selects the reaction tier. Three one-sentence lines synthesized via ElevenLabs eleven_turbo_v2_5, streamed as mp3_44100_128. The user hears a calm voice say what word-of-mouth actually sounds like. No dashboard metric converts skeptics faster.
Same deep-search engine, different framings. Because the question isn’t just “is this ad original?” — it’s “is this brief original?” and “is this claim defensible?” and “how do I get around the prior art we just found?”
Drop in headline, body copy, script, industry, media type. Get a 0–100 originality score with four dimension breakdowns, evidence citations, and a voiced reaction.
Drop a product claim (“50% fewer calories,” “engineered for professionals”) and find out if the category already says it. Same engine, different prompt.
Side-by-side comparison of two ads on all four dimensions with shared evidence. For A/B decisions in a pitch room.
Trend analysis over the cached corpus — what concepts, structures, and languages are spiking in a given industry right now.
Give Claude the brief and the scored evidence; get back three directions designed to avoid the prior art found in the search. Originality as a prompt, not an accident.
Effectiveness scoring against benchmarks — does this ad have the structural features that correlate with performance in its category?
Before the agency even writes the deck, score the brief itself. Catch the derivative ask before it turns into a derivative asset.
Every ad analyzed, archived. PDF report export via @react-pdf/renderer. Re-run with updates. Share with the team.
No custom ML. No proprietary corpus. No moat except the orchestration. Three days of spec-driven development with Kiro.
Kiro is an AI-native IDE built around spec-driven development — write the spec, generate the implementation, iterate. Ad-Visor 3000 was the test case: could you ship something genuinely useful to the advertising industry in a seventy-two hour window?
Three days. One person. No pre-existing ad corpus, no prior scaffolding beyond create-next-app. Spec-driven development with Kiro as the primary interface — Claude doing reasoning inside the IDE, humans shaping the spec.
A four-facet semantic decomposition pipeline via Claude. Multilingual embeddings and vector search. Firecrawl web expansion of the corpus on every query. Dimension-specific Claude judgment with cross-dimension inheritance rules. ElevenLabs voice reactions calibrated to score. PDF report generation. Eight functional routes. NextAuth authentication with bcrypt. Netlify deployment on the Next.js plugin.
Spec-driven development plus a strong primitives stack (Claude for reasoning, TurboPuffer for search, Firecrawl for web, ElevenLabs for voice) collapses what would be a month of engineering into three days — provided you’re willing to be ruthless about what the product actually is. Ad-Visor 3000 is not a dashboard, not a platform, not a marketplace. It answers one question, with receipts, out loud.
Evaluating agency decks. Catch the derivative before the invoice clears. Paste the headline, click analyze, get a citation list in sixty seconds.
Sanity-checking campaign briefs. Catch the derivative ask before it turns into a derivative asset. Intervene upstream via /brief-score.
Got burned before. Want evidence, not vibes. Ad-Visor 3000 hands them a citation list with URLs — “here are the eight ads yours is derivative of,” in order, with links.
Stress-testing their own lines before they go into a pitch deck. Nothing builds confidence like a 91/100 originality score on your own headline.
The good ones use it internally before presenting. The rest get found out. Either way: better work reaches the client.
Checking claims before they go on pack, on air, or on screen. Same deep-search engine, via /claims. “Who else says this, and how defensibly?”
Drop in a headline. Get a score, four dimension breakdowns, cited prior art with URLs, and three one-sentence reactions from the people your ad has to survive. Sixty seconds. Free to use.