Add one guard clause, in Chris Heckman's own words, to the Discover step of competitors-kb/miner-methodology.md and to any Claude run that asks "what do my winners have in common": "Do not give me feedback until you have verified that you are actually able to see the ad creative. Inferring from the naming convention alone is not enough." Operationally: the run must load the actual image files from competitors-kb/proven-winners/<theme>/cropped/ (they already exist, cropped to just the printed artwork) and print three visual attributes per design — line weight, palette, text-vs-graphic ratio — before it is allowed to state a pattern. Never let it reason from the title field in manifest.json, from Shopify handles, or from your Klaviyo naming convention. Run it first on the 57 files in proven-winners/garden/plant-humor/cropped/, not on faith-gardening — that is where the title pass scored zero, and it is the lane your own best-sellers live in.
Your competitor corpus was labeled twice, once from product titles and once by looking at the pictures, and both labels were kept — so it contains a measured proof that titles cannot see humor, which is exactly the failure Chris wrote a guard clause against on call #32.
Call #32's transcript addendum contains a prompt Chris dictated live for analysing six months of ad creative. Buried in it is a rule that reads like paranoia: don't give me feedback until you have verified you can actually see the creative; inferring from the naming convention is not enough. It sounds like prompt-engineering hygiene.
It is not, and IGL has the receipt. The proven-winners corpus was built twice — once classifying each design from its title, once by a vision pass over the image — and both labels survive in manifest.json. Of 620 designs, 333 (54%) got no usable label from the title at all, and 89 of the 287 that did were overturned. Title-only was silent or wrong on 422 of 620, 68%.
But the aggregate hides the real finding, which is that the error is directional, not random. Sorted by theme, the failure is not spread evenly: it concentrates almost perfectly in humor. Of the 57 designs the corpus finally classifies as plant-humor, the title pass identified zero. Not a low number — zero. It filed 26 of them as "general-garden" and 13 as "veg-homestead," and the 4 designs it did call plant-humor were all something else. The reason is plain once seen: a product title says what a shirt is about (gardening, vegetables, plants), while the joke lives in the picture and the phrasing. "Maybe Plants Are Addicted to Me" reads as general-garden to a title classifier. So does "Most Likely To Talk To Plants."
That matters more for IGL than for any other shop in the corpus, because humor is IGL's proven winning mode. The four best-selling shirts of the last 90 days — Listen to the Birds, Quietly Cheering for Moss, Life's Too Short Tomatoes, Excuse Me Bird Feeder's Empty — are all wry garden humor. The single category the mining pipeline is blindest to when it reasons from titles is the exact category IGL wins in.
Why no single shelf could produce this: the mastermind call carries the rule but no proof; the corpus carries the proof but never states the rule; and the Wright Brothers episode carries the reason the rule matters at all. Lilienthal's flight tables were trusted by the entire field and were, in Orville's word, worthless, so the brothers built a wind tunnel and measured for themselves. A product title is Lilienthal's table: an authoritative-looking label somebody else wrote, standing in for a measurement nobody took. Dave Shaffer's Claude run that surfaced woodcut/linocut across his top 8 designs (call #27) is the counter-case — that answer is only reachable by looking at pixels, since no product title in his store contains the word "woodcut."
wescale-kb/calls/call-32 — SEED. Transcript addendum, Segment 3: Chris Heckman's live-dictated creative-analysis prompt, carrying the verbatim guard clause at line 229. Raw, unenriched call transcript.competitors-kb/proven-winners/manifest.json — the measured proof. 620 records, each carrying theme_title (title-derived), theme_vision (image-derived) and theme_final. Corpus built 2026-07-12. This is a classification artifact, not sales data — the knowledge bases hold no product-level sales figures.founders-kb/episodes/028-the-wright-brothers — the principle underneath the rule: "Clearly, those esteemed authorities had been guessing, groping in the dark. The accepted tables were, in a word, worthless. We had to go ahead and discover everything ourselves" (Orville Wright, line 93). They built a wind tunnel rather than trust Lilienthal's published tables.wescale-kb/calls/call-27 — the counter-case: Dave Shaffer had Claude analyse his top 8 designs and it surfaced woodcut/linocut engraving as the shared trait, a finding only reachable from the images.competitors-kb/proven-winners/README — theme aggregate. Documents the two-pass method that makes the disagreement measurable, and confirms every theme folder has a cropped/ subfolder of artwork-only images ready to feed the analysis.competitors-kb/miner-methodology.md and to any Claude run that asks "what do my winners have in common": "Do not give me feedback until you have verified that you are actually able to see the ad creative. Inferring from the naming convention alone is not enough." Operationally: the run must load the actual image files from competitors-kb/proven-winners/<theme>/cropped/ (they already exist, cropped to just the printed artwork) and print three visual attributes per design — line weight, palette, text-vs-graphic ratio — before it is allowed to state a pattern. Never let it reason from the title field in manifest.json, from Shopify handles, or from your Klaviyo naming convention. Run it first on the 57 files in proven-winners/garden/plant-humor/cropped/, not on faith-gardening — that is where the title pass scored zero, and it is the lane your own best-sellers live in. ## Why (the one-liner) Your competitor corpus was labeled twice, once from product titles and once by looking at the pictures, and both labels were kept — so it contains a measured proof that titles cannot see humor, which is exactly the failure Chris wrote a guard clause against on call #32. ## The insight Call #32's transcript addendum contains a prompt Chris dictated live for analysing six months of ad creative. Buried in it is a rule that reads like paranoia: don't give me feedback until you have verified you can actually see the creative; inferring from the naming convention is not enough. It sounds like prompt-engineering hygiene. It is not, and IGL has the receipt. The proven-winners corpus was built twice — once classifying each design from its title, once by a vision pass over the image — and both labels survive in manifest.json. Of 620 designs, 333 (54%) got no usable label from the title at all, and 89 of the 287 that did were overturned. Title-only was silent or wrong on 422 of 620, 68%. But the aggregate hides the real finding, which is that the error is directional, not random. Sorted by theme, the failure is not spread evenly: it concentrates almost perfectly in humor. Of the 57 designs the corpus finally classifies as plant-humor, the title pass identified zero. Not a low number — zero. It filed 26 of them as "general-garden" and 13 as "veg-homestead," and the 4 designs it did call plant-humor were all something else. The reason is plain once seen: a product title says what a shirt is about (gardening, vegetables, plants), while the joke lives in the picture and the phrasing. "Maybe Plants Are Addicted to Me" reads as general-garden to a title classifier. So does "Most Likely To Talk To Plants." That matters more for IGL than for any other shop in the corpus, because humor is IGL's proven winning mode. The four best-selling shirts of the last 90 days — Listen to the Birds, Quietly Cheering for Moss, Life's Too Short Tomatoes, Excuse Me Bird Feeder's Empty — are all wry garden humor. The single category the mining pipeline is blindest to when it reasons from titles is the exact category IGL wins in. Why no single shelf could produce this: the mastermind call carries the rule but no proof; the corpus carries the proof but never states the rule; and the Wright Brothers episode carries the reason the rule matters at all. Lilienthal's flight tables were trusted by the entire field and were, in Orville's word, worthless, so the brothers built a wind tunnel and measured for themselves. A product title is Lilienthal's table: an authoritative-looking label somebody else wrote, standing in for a measurement nobody took. Dave Shaffer's Claude run that surfaced woodcut/linocut across his top 8 designs (call #27) is the counter-case — that answer is only reachable by looking at pixels, since no product title in his store contains the word "woodcut." ## Evidence trail - wescale-kb/calls/call-32 — SEED. Transcript addendum, Segment 3: Chris Heckman's live-dictated creative-analysis prompt, carrying the verbatim guard clause at line 229. Raw, unenriched call transcript. - competitors-kb/proven-winners/manifest.json — the measured proof. 620 records, each carrying theme_title (title-derived), theme_vision (image-derived) and theme_final. Corpus built 2026-07-12. This is a classification artifact, not sales data — the knowledge bases hold no product-level sales figures. - founders-kb/episodes/028-the-wright-brothers — the principle underneath the rule: "Clearly, those esteemed authorities had been guessing, groping in the dark. The accepted tables were, in a word, worthless. We had to go ahead and discover everything ourselves" (Orville Wright, line 93). They built a wind tunnel rather than trust Lilienthal's published tables. - wescale-kb/calls/call-27 — the counter-case: Dave Shaffer had Claude analyse his top 8 designs and it surfaced woodcut/linocut engraving as the shared trait, a finding only reachable from the images. - competitors-kb/proven-winners/README — theme aggregate. Documents the two-pass method that makes the disagreement measurable, and confirms every theme folder has a cropped/ subfolder of artwork-only images ready to feed the analysis. ## Worked examples — the FULL result set, all 17 theme folders Rate = share of that folder's designs the title pass got wrong or could not label at all. | Rate | n | blank | overturned | theme (folder) | |---|---|---|---|---| | 100.0% | 57 | 12 | 45 | plant-humor — start here | | 100.0% | 28 | 27 | 1 | horse-heritage | | 100.0% | 23 | 21 | 2 | misc-nongarden | | 100.0% | 2 | 2 | 0 | slow-living | | 88.9% | 18 | 15 | 1 | spiritual-empower | | 88.2% | 17 | 15 | 0 | outdoor-adventure | | 86.2% | 29 | 23 | 2 | faith-gardening | | 82.4% | 17 | 10 | 4 | moss-mushroom | | 77.8% | 36 | 25 | 3 | food-italian | | 74.6% | 71 | 44 | 9 | garden-wildlife | | 67.2% | 58 | 36 | 3 | nature-outdoor | | 61.5% | 26 | 16 | 0 | auto-vintage | | 53.3% | 60 | 25 | 7 | veg-homestead | | 46.9% | 32 | 13 | 2 | general-garden | | 42.9% | 77 | 27 | 6 | wildflowers-floral | | 41.9% | 43 | 15 | 3 | pollinators | | 30.8% | 26 | 7 | 1 | ocean-coastal | Individual designs the title pass misfiled, all verified in manifest.json (brand-safe examples only): - "Maybe Plants Are Addicted to Me" — title said general-garden. Actually plant-humor. - "Most Likely To Talk To Plants" — title said general-garden. Actually plant-humor. - "Oops, I Bought More Plants" — title said general-garden. Actually plant-humor. - "Botanically Unhinged" — title said wildflowers-floral. Actually plant-humor. - "Certified Overplanter" — title said general-garden. Actually plant-humor. - "WTF: What's That Flower?" — title said wildflowers-floral. Actually plant-humor. - "Floral Heart T-Shirt" (faithwiseco.com) — title said wildflowers-floral. Actually faith-gardening, IGL's own core niche. - "In The Garden T-Shirt" (gardenistee.com) — title said general-garden. Actually faith-gardening. - "Ladybug Flowers T-shirt" — title said pollinators. Actually garden-wildlife. ## Assay verdicts - Novelty — PASS (4/5). The guard clause appears only in call-32.md:229 and its two transcript copies, purely as prompt hygiene for ad review; nothing ports it to mining, and nothing anywhere states the corpus's own title-vs-vision disagreement. Vault checked: #005, #010 and #031 all interrogate the same manifest and none touch classification provenance. Operator re-check: title-vs-vision returns zero hits across all 13 bases and all 32 prior nuggets. - Grounding — PASS (5/5). Every number reproduced independently from disk by both the judge and the operator: 620 records, 333 blank, 89 overturned, 422/620 = 68.06%. All worked examples verify verbatim. The guard clause is verbatim at line 229; the Orville Wright quote verbatim at line 93 of episode 028. - Actionability — PASS (4/5). Changes a named, repeated decision — how the Discover step classifies designs. Specific guard-clause insertion into a named file plus a hand-run over a named existing folder. No staff, capital, or inventory. Not a 5 because the payoff is analysis quality, one step removed from revenue. Thread: call-32 transcript addendum (Chris's "verify you can see the creative") → call-27 (Dave's top-8 woodcut finding, image-only) → founders ep-028 Wright Brothers (Lilienthal's worthless tables) → competitors-kb proven-winners/manifest.json (counted title-vs-vision labels: 68% wrong or blank, and zero of 57 for humor)