Adrien's feedback — every review, chronologically, and the rule each one created

Sources: /root/workspace/wh-remix/feedback/loom-b2.txt … loom-b9.txt (run 1, reviews of batches 02–09 = "Loom 2…9"), loom-s1.txt … loom-s4.txt (Sam batch = "Sam Loom 1…4"). The .mp4 sits next to each .txt. Transcripts are auto-generated from voice: numbers are sometimes misheard ("ATS 2" = #2, "S32" said for S33, "S66" for S56). Quotes below are short excerpts.

Two reviews have no transcript:

• Loom 1 = review of batch 01 (79 ads), given before the transcript habit. Its content survives in the "Feedback log — batch 01" table on https://blueprint.machineroom.app/protocols (reproduced in §1).
• Sam Loom 5 = S11 validation only ("made like theirs", A25) — recorded in validated.json.

---

1. Run 1 — the 9 rounds (79 templates, 0 % first try → 79/79)

Loom 1 (batch 01 review, from the protocols log)

Verdict: nothing validated as is. Batch 01 had been forced onto the research headlines (R9) and a context line (R10).

• 3 benefits were product features → B3. #4 added bottom line repeated the ad → B1. #5 too much text under it → B1. #6 extra text squeezed the bottom bar → B1. #7 didn't say WHO it's for by their problem ("men over 40" is not a problem) → B2. #8 problem not clear + extra line → B18/B1. #9 negative lines not precise → E8. #11 said "conversion" → B4. #14 changed the text although same angle → A6. #16 their headline was better (verb + body part + deadline) → A6. #17 angle problem — "Drains beer belly in 14 days" word for word → A6. #18 text not adapted at all → "Beer belly? Try Wildhorn." → A6. #25 "your 12 supplements" nobody gets it → B17. #29 customer problem became a product problem → B19. #57 concept sexual, adaptation wasn't → A4. #63 great template badly adapted → A1. All ads: too many "supplements failed you" → A5 (favour beer belly, man boobs).

• Taste signal: he wants the template's own words and mechanism, not the research headline pasted on top.

Loom 2 — batch 02 (#1–#20 redone)

Mood: first real progress ("9 is exceptional… really better than 10 now").

• Liked: #9 wojak/chad (exceptional), #13 #14 (well adapted), #16 (better), #18 (perfect), #19 (excellent), #20 (good), #11 angle "not bad".
• Disliked: #10 lost the Mars Men illustrated style; #11 twist less logical for a belly; #16 jar crushed.
• Demands: explain why #9 worked; understand why #16's generation failed; apply #17's fix to other ads; redo the bad ones + generate #20–30; always show the first generation next to the new one; feedback loop on good AND bad.
• Rules born: A27, C5, C6, D3, D4, D11, D12, E8.

Loom 3 — batch 03 (#1–#30)

• Liked: #6 (great work), #10, #11 (more), #16 (better), #21 (really great — "understand properly why"), #22 (amazing, "such a great work"), #24, #25 ("really upgrading the game… apply it everywhere"), #26 idea, #27 #28 #29 (better), #29 "idea way better".
• Disliked: #2 changed the side, body without arm ugly, logo wrong ("the first one was way better"); #8 man-boobs angle not hit; #23 text doesn't match, not visibly a beer belly, competitor's lab logo replaced by our product photo ("I'm really deceived"); #26 added hair to a non-hair product ("big feedback loop").
• Demands: QC that everything fits the angle; show all batches when making another version; another version of #30 using all data; keep adding every improvement; redo bad ones; do #30–40.
• Rules born: A2, C1, C4, C7, C8, C13, B11, B17, B18, B19, A28, D2, D3, E9, E10.

Loom 4 — batch 04 (#1–#40)

• Liked: #8 (amazing — "deeply into the avatar and the angle"), #23 ("perfect", text "really good"), #26 ("you really understood"), #29 (great improvement), #30 #31 #32 #34 (great, "adapted the angle to the product"), #36 (great), #37 (adapted the template perfectly).
• Disliked: #2 drew an ARM while we sell belly/man boobs; #16 forgot the little orange glow on the result; #23 "you should have removed the Lab — it's the competitor's brand"; #35 organ chunks "a bit disgusting"; #36 brisket in the background; #37 looks like a killed bison / bison poop, "we are not selling colostrum"; #38 vague rival + vague problem; #39 chocolate ("don't put chocolate in our product when it's not chocolate"); "nobody cares about a capsule… show the bison organs inside the capsule… think ahead visually".
• Rules born: A2 (the Template DNA step — biggest single rule of run 1), A28, B5, B18, C2, C4, C14, D11.

Loom 5 — batch 05

• Liked: #2 (last one great), #16 (great), #23 (perfect), #36 (really smart), #37 (great — "the mechanism has to be different than the example, but still the same idea"), #38 (great), #40 ("way bigger improvement"), #41 #42 (perfect), #44 (makes sense), #46 #49 (great), #47 (great image).
• Disliked: #35 still not instantly "bison organs"; #38 bison hard to recognise; #39 still not visual enough (want heart, liver falling into the bottle + capsules); #43 "one notch tighter" not understandable; #44 doesn't like the example; #47 "kind of weird" → completely different version.
• Rules born: A29, B9, C2, C8, C15, C16.

Loom 6 — batch 06

• Liked: #35 B6 more understandable; #39 ("I love it… exactly what I was thinking"); #44 B6 ("even better than the example"); #47 B6 ("so aggressive… the greatest"); #51 "50% off for life" wording; #52 #53 (smart comparison) #55 #56 #59 #60 (great); #57 ("really hard ad, you got it pretty well").
• Disliked: #35 B6 lost the winner's principle (billboard made of the product); #51 mentioned Autoship ("we never speak about auto-shipping"); #54 avatar doesn't wear compression shirts (validated anyway); #58 same angle → the bottle itself should get fat ("their ad is better… don't touch the product design").
• Demands: understand #39's difference maker "again and again"; #57 in a beer-belly version.
• Insight: "my customer is not that smart" — simplest, most concrete proof wins.
• Rules born: A6 (visual mechanism), A14 (seed), B7, B8, B9, B10, B15, C3, C6, E11, E12.

Loom 7 — batch 07

• Liked: #35 B7 ("you really understood what was the game"), #51, #58 ("great, we're getting there"), #64 B7, #66, #67, #68 (great), #69 removing the pineapple, #70 capsule of organs ("exactly what I imagined").
• Disliked: #57 B1/B7 not understandable (horn not sexual enough; tailor not readable as weight loss); #61 missed the vulva innuendo; #62 changed the design because GPT refused → "go for another model… you have the API key"; "the ads you made are way not aggressive enough"; #63 failed "let no one hit"; #65 "shirt off since day 21" unreadable + invented timeline; #68 "belly" not "beer belly"; #69 sexual line, belly visual; #70 shower copied for no reason.
• Rules born: A3, A4, A6, B2, B6, B10, D1, E3.

Loom 8 — batch 08

• Liked: #57 woman's hand idea; "I love how you're speaking to her"; #62 #63 (amazing), #65 line "melt his beer belly before you drain him" (great), #66–68, #70 adaptation, #72, #74 (great), #75 (great but…), #78 (perfect), #79 angle.
• Disliked: #57 jar not properly in her hand + added text; #61 kept their pineapple ("it's a vaginal probiotic"); template style not respected; #65 deadline added, jar not on the sofa/lounger, flat text instead of sunscreen bubbles; #69 still needs something more sexual that fits product + end goal; #70 B7 capsule was better, feet should be on the scale; #71 organs look like meat; #73 text not centred on the eggplant; #74 → make it beer belly; #75 kids' beds kept, need beer-belly elements; #76 "organs, organs, organs"; #77 not the same body; #79 older man, jar in front of the face, add a woman's name.
• Rules born: A7, A8, A9, A10, A11, A12, B6, B12, B13, C2, C5, C9, E1, E2.

Loom 9 — batch 09

• Liked: #57 ("great… perfect"), #61, #65 idea, #71 (validated), #72 B9 ("amazing"), #76, #79 ("perfect… the older guy").
• Disliked: #57 product too small; #65 template quality higher (3D product on the table, text texture); #69 "better but still doesn't make any sense"; #70 organs disgusting — no veins; #73 small text too close to the headline ("failed again"); #77 navel not perfectly aligned (better).
• Verdict override: #74 keep B8.
• Demand (the most important of run 1): "#75 — you could have done this from the first try… every pixel of the scene, then understanding what you have to change for the customer avatar… make a deep analysis of every ad." → A13.
• Rules born: A13, A9, C2, C9, C10, C17, A26, E4.

---

2. Sam batch — 4 rounds (83 templates, 71 % first try → 83/83)

Sam Loom 1 — S1–S10 (6/10 first try)

• Liked: S1, S3–S6 ("great"), S8 (great), S10 (great).
• Disliked: S2 not on the beer belly; S7 not shocking enough ("the belly melting or something more shocking"); S9 not understandable for a stranger.
• Fix: S8 silhouettes face the result, left → right. Question: S10 "why sexual and not beer belly?"
• Rules born: A14, A15, A16, B14, E5.

Sam Loom 2 — S2 S7 S8 S9 redos + S11–S40

• Liked: S7 both tries (count as success), S9 both tries ("try one was great"), S12, S13 scene, S16 (perfect), S19–S22, S23 ("really great understanding"), S25–S28 ("I like your way you adapted that"), S30, S31, S32 ("so aggressive, I like it"), S34, S36–S38; S33 calorie-app idea.
• Disliked: S8 first silhouette not aggressive enough (arm, push the belly right); S11 jar not on her head ("easy to make… explain why"); S14 should have copied; S15 shoelaces could sell a shoelace brand ("really bad"); S17 brown organs ("look at their organs… you want to eat them"); S18 should pour beer into the beer ("lack of understanding of the whole concept"); S24 stat must be a benefit claim; S29 simpler words; S33 phone badly held + bed copied from a sleep ad; S35 off-angle.
• Meta: "I want everything I tell you to be an improvement." "A lot of improvement, but some things are really wrong."
• Rules born: A15, A17, A18, A19, A3, A14, B15, B16, C2, C11, C18, D6, D11, E6, E7, E13.

Sam Loom 3 — round-2 redos + S41–S83 (31/43 new on first try)

• Liked: S15 ("adapted perfectly to the problem"), S17, S24 ("you put exactly what I told you"), S33 try 2, S35 try 2, S42 (said "amazing" then hesitated), S44–S49, S51, S53, S54, S55 text ("perfect"), S58 try 1, S59–S73 (S67 "really good understanding of the joke"), S75–S78, S79 try 1, S81–S83.
• Disliked: S11 the girl's face IS the product, painted style ("you fucked it up… maybe go out of GPT"); S18 powder makes no sense (capsules) + logo failed twice, try 1 better; S29 try 3 changed the style; S35 logo again; S42 must say it's for the beer belly ("you didn't respect the formula"); S50 sexual fruit → fat fruit vs lean fruit; S52 off-angle ("why sometimes you get out of the angle I told you, which is beer belly"); S55 the woman should be on top; S56 reads like a hangover product; S57 → woman angle tied to the belly; S58 why make a try 2; S74 two lean bellies instead of fat vs lean; S79 highlight only man boobs; S80 not understandable.
• Rules born: A20, A21, A22, A23, A24, A30, A31, A8, D8, D13 (audit), E13.

Sam Loom 4 — the last 12

• Liked: S14, S18 ("great job for the last one for the beer"), S29, S42 try 1 ("I validate it… do not change"), S50, S52, S56 (background doesn't make sense but it goes), S57 ("you adapted the angle properly"), S62, S74, S80.
• Disliked: S11 still failing — "review the ad by screenshot and understand the model they use; it's not realistic; it succeeds, so why can't you?"
• Rules born: A25, A26, A3 (setting must fit the claim).

Sam Loom 5 (no transcript)

S11 v5 validated: their photo edited minimally by Seedream (aged man, text cleared, 9:16), jar swapped with jarswap.py, text composited. → 83/83.

---

3. Loom coverage matrix — every distinct piece of feedback → rule

Rule IDs refer to wildhorn/lessons-wildhorn.md. "E" = repeat-what-worked rule. Rows marked † are positive verdicts whose lesson is the E/A rule cited. Total: 261 feedback items, 0 uncovered. 23 rules were added by this audit (A27–A31, B17–B19, C12–C18, D11–D13, E8–E13; D13 comes from the file audit, the other 22 from Looms).

loom-b2.txt (Loom 2) — 15 items

loom-b3.txt (Loom 3) — 25 items

loom-b4.txt (Loom 4) — 26 items

loom-b5.txt (Loom 5) — 20 items

loom-b6.txt (Loom 6) — 22 items

loom-b7.txt (Loom 7) — 28 items

loom-b8.txt (Loom 8) — 30 items

loom-b9.txt (Loom 9) — 17 items

loom-s1.txt (Sam Loom 1) — 7 items

loom-s2.txt (Sam Loom 2) — 23 items

loom-s3.txt (Sam Loom 3) — 37 items (S3-1…S3-37)

loom-s4.txt (Sam Loom 4) — 11 items (S4-1…S4-11)

Count: loom-b2 15 + b3 25 + b4 26 + b5 20 + b6 22 + b7 28 (rows 109–136) + b8 30 + b9 17 + s1 7 + s2 23 + s3 37 + s4 11 = 261 feedback rows, every one mapped. (The run-1 + Sam Loom 1–2 rows are numbered 1–213 continuously; Sam Loom 3–4 rows use the S3-/S4- prefix.)

---

4. How Adrien reviews — his preferences, distilled

Format. He records a Loom (≈2–17 min), scrolling the review page ad by ad, thinking aloud in English (French accent, auto-transcribed). He names ads by number ("the 23", "S42"). He judges in one look; anything he must decode, he calls "not understandable".

What makes him say "great / amazing / perfect":

1. The template copied almost pixel for pixel, with only the nouns swapped for our world (#21, #25, #63, S14, S83). His mantra: "you should have just copied the ad".
2. Deep avatar precision — his shame, his routine, his proof (#8 hide-the-chest, #9 wojak scenes, #44 old jeans, #47 lights on).
3. Aggression. "Way not aggressive enough, so it's shit." Sexual templates stay sexual; shocking templates stay shocking.
4. Organs he can SEE and would want to eat. Never a generic capsule, never meat chunks, never veins, never brown.
5. The beer belly named everywhere. He gets visibly annoyed when an ad drifts to brain, hangover, muscle, 30s, generic wins ("I don't know why sometimes you get out of the angle I told you").
6. Visual quality equal to the template: 3D product with real shadow, textured text, same render style, correct text spacing.

What makes him angry ("you fucked that up"):

• Importing their product world (pineapple for a vaginal probiotic, chocolate, hair, brisket, arm, their lab logo).
• Changing something he already validated (S18, S58, S79, #2) or changing the style when he asked for wording (S29).
• The same mistake twice (logo misspelled again, text spacing failed again, S11 three times).
• Not explaining WHY: every Loom ends with "feedback loop — find why you failed, never do it again".

His stable preferences:

• Same amount of text as the template; no added line, tagline, disclaimer, deadline.
• Name the problem (beer belly), never the age. Simple words — "my customer is not that smart".
• Our avatar everywhere: 45–60, greying, beer belly; women in sexual ads are the partner.
• Offer framing without subscription words ("50% off for life" is fine, "Autoship" is not).
• A customer-review template keeps a first name + initial.
• His latest verdict wins; "try 1 was better" = freeze try 1.
• He wants to see progress: template + first version + new version side by side, every round, and a hit-rate comparison.
• He wants the system to learn: every correction becomes a permanent rule (LESSONS + lint), and he expects the NEXT batch to be right "from the first try".
• Working language for this material: English (site, prompts, briefs, replies).