🏆 Gold Nuggets · September 27, 2026

Luca Ferrari did not trim job titles at Bending Spoons, he abolished them. He traced the practice to its root benefit, found the benefit belonged to the employee's CV rather than to the company, and removed the whole thing: "pick your own, we don't need to know, just don't embarrass us." His general rule is that incremental simplification gets you 5% while removal can get 10x, and that if 2% of users touch a feature you cut it whole rather than running a million migrations. The transferable move is not deletion for its own sake. It is deciding a category question once, explicitly, instead of letting a process answer it by default every cycle.

IGL's occult lane is exactly that unanswered question. The brand has a real, current, written exclusion — igl-no-dark-humor.md, marked Status: CURRENT canon, Tim's directive of 2026-07-09. It is specific where it speaks: skulls, poison, "I Bury Things", Feral, rage. It is silent on witches, spells and wizards. And its enforcement clause aims at the mining board, not the publish step. So the rule is not weak, it is narrow and early — and a batch that arrived by a new route found nothing in its way.

The non-obvious part is that the sales figures cannot carry this argument, and that is the finding too. All four shirts have zero net units — but so do 45 of the 48 listings in their own batch, and 627 of the 1,318 live listings overall. Their whole batch produced 3 units. "They didn't sell" is the least interesting true thing about them; almost nothing published that week sold. What distinguishes them is not performance, it is permission: they are the only genuinely occult titles in a 1,318-item catalogue belonging to a faith brand, and they all came from one batch on one day. This is not a habit to break. It is a single unguarded gate, which is a much cheaper thing to fix and a much easier thing to miss.

No single shelf holds this. founders-kb supplies the discipline of settling a category once rather than per-cycle. Only the Shopify capture in wescale-kb can say which titles are live, when they published, and that the sales data refuses to distinguish them. Only igl-brand-kb holds both halves of the judgment: the written ban with the occult-shaped hole in it, and Lydia the Green Witch, who is the reason the fix has to be about the merchandise and not the audience.

Musk's step 2 is delete the part, and his test is whether anyone can name who required it (founders-kb ep-399, Key Lesson 1: the battery-pack caps were deleted once nobody could name who required them). IGL runs the inverse. 81 of 312 summer listings — 26% — are duplicate-title siblings in 28 families, and nobody can name who required eight copies of Red, Ripe & True.

The cost is not storage; print-on-demand listings are free to hold. The cost is that every report and the catalog algorithm judge a listing, while customers buy a message. Red, Ripe & True is the clearest case: 17 of the summer's 192 total first-42-day paid orders — 8.9% of everything sold — carried by a message that never appears at the top of any list.

Where splitting provably changed a decision, stated narrowly. The cohort's documented graduation rule is Jakob & Petra's, from call-28: run a low-budget tag-based catalog on every new launch, then after a week pull the designs with 10+ clicks into static tests. Measured against that bar across all 28 duplicate families, splitting flips the decision for three: spirit-of-america-raccoon (14 clicks combined, best single 6), tomato-quilt-club (11 combined, best single 9), and Red, Ripe & True itself — 25 clicks combined, but only 1 of its 8 listings clears 10 on its own, and the second-best sits at 9, one click short. The other 25 families are unaffected. The splitting is systematic; its measured effect on graduation is not, and saying otherwise would overstate it.

The loop back to the brand's own shelf is the tell. IGL's email C169, "The story behind Red, Ripe and True," went to the 60-Day Engaged segment on 2026-07-01 — six days after the eight listings launched. Six of the seventeen orders landed on listings with zero catalog clicks. The story email found the winner; the ad algorithm, looking at eight small pieces, never did.

No single shelf reaches this: founders-kb supplies the deletion test, wescale-kb supplies both the 312-listing artifact and the graduation rule, igl-brand-kb supplies the email that actually sold it.

The seed was a Skool lesson about turning a video into an ebook, and its real subject is not publishing — it is the fact-check step. Every claim from the draft gets a verdict and evidence checked against a corpus the author did not use, and that report, not the article, is the single human checkpoint: "You give it a URL. You read the fact-check report. You approve." (agentic-academy-kb/lessons/00-youtube-to-ebook.md, line 105.)

IGL's persona dossiers are exactly that draft. 02-donna-seed-keeper.md declares status: CURRENT SoR and lists its evidence in the header — ad comments, community stories, Facebook voice, a sales cluster. The review export is not on that list. So when the file reached the question of who wears the shirt, it had a handful of quotes and honestly wrote the question down as open rather than guessing. Run the same claim against the corpus it skipped and the answer arrives with a size distribution attached, which quotes alone could never produce: 78% of man-as-wearer reviews are L/XL/2XL, against a store-wide top three of L (229), XL (187), M (134).

The sharpest single line in the corpus is a one-star review on Quietly Plotting My Next Tomato Harvest: "the shirt so short it was barely long enough for me, a woman just over five feet tall. I wanted the shirt for my six foot son." Nugget #022 established that IGL's own reviews complain shirts run too long; this is the same garment failing in the opposite direction for the buyer this nugget is about. Both are true, and they are two different people.

No single shelf could reach this. The Skool lesson supplies the method (claim → verdict → evidence, against a corpus the author skipped). The brand base supplies the claim and the flag that it was unresolved. The mastermind base supplies why concentration matters commercially: Antonio Rodriguez's profile names "over-concentration of ad spend on a single design" as the root cause that took his week down (line 76) — and 72 of the 103 man-as-wearer reviews sit on the bird-feeder family, the same design that is already 364 of all 921 reviews (39.5%). He also ran a "Who are you shopping for?" popup before Black Friday 2025 where roughly 50% answered "for someone else" (line 85).

Honest limit: these are review counts, not sales. The thirteen knowledge bases hold no product-level sales data, so this measures who shows up in review text, not units shipped. The corpus itself sits outside the knowledge bases, in Tim's Drive folder.

Zach Dell's operating doctrine has one load-bearing rule, and Base Power hard-codes it into an object. Antonio Gracias opens every conversation the same way — "Zach, what's your constraint? Like, what's the constraint of the business? And I'll just answer the question, and that's all we can talk about" — for years. Then there is a literal physical turtle that "goes on the desk of the person that is on critical path," and "that turtle can only be moved when critical path is resolved," with a hot-potato daily meeting and an explicit resolved email at the end. Dell names the failure mode precisely: "it's kind of like a cope, if you will, to talk about the other stuff."

IGL already has the diagnosis. The coach gave it twice, in the scorecard database's scorecard_reviews table. At call #31 Tim's own stated number one was conversion rate and Chris Heckman overruled him: "I would redirect from CVR. The number one is the margin leak, and until that's fixed, nothing else tells us the truth... that's not the constraint this week." At call #34, on $1,088.89 of Facebook spend: "The goal right now is to stop the bleed, not add fuel to it. Next week I want to see you profitable on whatever spend you have in market, even if it's small."

What happened next is structural rather than lazy. Across the call files carrying a named IGL-application heading from call #34 onward, the weekly routine generated 41 fresh IGL action items. Over the same span net margin was negative in every reported week — scorecard weeks 40 through 47 run −39.6%, −32.2%, −41.0%, −44.0%, −33.2%, −51.0%, −29.0%, −51.3% — Facebook return on ad spend never once exceeded 1.00, and spend went from the $1,088.89 Chris was reacting to up to $2,569, about 2.4 times the number that triggered "stop the bleed."

No single knowledge base could produce this. The doctrine and its enforcement mechanism are in one founder interview; the coach's verbatim diagnosis and its date are in a database table; the 41 competing instructions are scattered across ten separate call summaries; and Tim's own recall-kb capture of Jon McNeill supplies the same rule from the Tesla side — the fix came from noticing the one bottleneck, not from working a list.

WeScale call #34 floats a comment-to-design pipeline as a concept — scrape the Facebook comment requests, feed them back through the design pipeline, repost — and Chris tells the room to validate it by hand first, because he found thousands of requests in his own archive (wescale-kb/calls/call-34.md:39). IGL's archive is already pulled and sitting in a JSON file. It has been mined twice — once for voice, once for drawing defects (nugget #035) — and never once for the thing Chris actually described.

Run it over the whole corpus and the concentration is extreme. Of 1,722 ad posts, only 196 ever received a comment, and 1,252 comments — 39.0% of every comment IGL's ads have ever collected — sit on ten posts for one design, the bird feeder shirt. Hand-classify every comment naming an animal and the request order is raccoon 7, deer 5, bear 1. (On raw mentions bear 9 edges deer 8, but seven of those nine bears are people telling bear stories, not asking for a shirt. Requests are the honest unit, and on requests the order holds.)

Here is the part no judge reached, and the reason this is not "make more animal variants." IGL already sells nine raccoon shirts. Across all nine, lifetime: 37 units, $1,166.79 — 0.33% of the store. Six of them, every one created in 2026, have sold zero. And six times between January and June 2025, a request for a raccoon was answered in-thread with a link to one of those existing raccoon tees — four of them within a day of the ask. The asks kept coming after each reply. The existing raccoon shirt is not a near-miss on the request; it is an answer to a different question. The bird feeder family, over the same measured period, is 4,322 units and $137,613.52 — 38.8% of all store units.

That distinction is where the two sources collide. The mastermind's answer to a slow week is breadth — call #34 line 38, "100 designs/week is not too many." Todd Graves built Raising Cane's on the opposite instinct and is explicit about the pressure he refused: the chains around him kept adding menu items to dodge the "veto vote," he studied In-N-Out ("since 1948, they have the exact same menu") and concluded "you can do one thing and do it better than anybody else." He draws the line that matters here too — "it's not simple. It's focused." A raccoon version of the bird feeder shirt is In-N-Out's animal style: a variation on the one item people already come for. Nine unrelated raccoon shirts are the veto-vote menu.

No single shelf produces this. The mastermind supplies the mechanism and the instruction to hand-validate it. A 2025 restaurant interview supplies the discipline that says spend the variation on the proven item. Only IGL's own comment archive says which animal. And only the sales database says the obvious objection — "we already have raccoon shirts" — is exactly the mistake.