The feedback loop — from a Loom to a permanent rule

The loop is why the Sam batch hit 71 % first-try validation after run 1's 0 %. Every review becomes rules; rules become lint checks where possible; the next batch is generated against all of them. The rulebook only grows.

review page ──► reviewer records a Loom (ad by ad, says the S number)
     ▲                         │
     │                         ▼
 regenerate ◄── rules.md + lint.py ◄── verdict table (WIN → why → repeat · MISS → root cause → rule)
 (redo + 10 new)                        │
                                        ▼
                         validated.json · sell_override.json · protocols log · stats

---

1. Ingest the Loom

cd $SAC_WORKDIR
bash /root/claude-vault/skills/static-ads/scripts/ingest_loom.sh "<loom url>" s5     # yt-dlp + mr-transcribe (OpenAI API)

• Transcription = mr-transcribe (OpenAI whisper-1 API, ~$0.006/min). Never local whisper on the VPS (it saturated the server for 3 days in Sept 2026).
• If the reviewer pastes the video file or a transcript, save it as feedback/loom-<id>.txt the same way.
• After transcription, the mp4 is heavy media: archive it (mr-archive push) or delete it once the transcript is verified; the .txt stays.
• Read the transcript in full, in the main context (never delegate the reading of a review: nuance is the whole point).
• Speech-to-text garbles ad numbers and brand words ("Pierre Bailey" = beer belly, "Manboub" = man boobs, "S66" said for S56, "67" said for 57). Resolve every number against the review page order and the ad content before writing a verdict.

2. Build the verdict table (one row per ad mentioned)

Verdict vocabulary (it feeds validated.json and the stats):

• ✅ Validated (first try) — validated on the first review of the first try.
• ✅ Validated (redo 1) / ✅ Validated (redo) / ✅ Validated (both tries) / ✅ Validated (try 1 = sell) — validated at the 2nd review.
• ✅ Loom N (after fix) / ✅ Loom N — try 1 validated, keep it — validated at the 3rd review.
• ❌ <reason> — to redo. ⏳ To review — not reviewed yet.
• Ads he does not mention: ask, or keep ⏳. Never assume a silent ad is validated. (S62 was never mentioned until Sam Loom 4.)
• If an ad needed more than 3 reviews, write the round count to rounds.json ({"11": 4}) so stats stay exact.

3. For every WIN: why → repeat

Write the mechanism that made it work, in one sentence a future batch can apply to another template. If it is new, add an E rule to 3-improve/rules.md (E1–E13 exist). A win without a "why" is lost knowledge: the reviewer explicitly asks for it ("feedback loop what you did on the 6", "find what you did right and try to do it more").

4. For every MISS: root cause → permanent rule → lint check if mechanizable

1. Root cause, not symptom. "Brown organs" is the symptom; the root cause was over-correcting an earlier rule ("not gross") into dried shapes. "Jar not on her head" was a symptom; the root cause was a fallback that silently dropped a placement (→ D6 parity check).
2. Rule. Phrase it brand-agnostically with the Wildhorn case as evidence. Extend an existing rule if it is a refinement (C2 was rewritten when S17 contradicted it — the history line stays), otherwise a new ID continuing the series.
3. Lint. If a regex or a count can catch it, add it to scripts/lint.py (generic) or to brand.json banned / visual_checks (brand-specific). Examples that became lint: D6 jar count, D7 seedream_full, D8 bare wordmark line, D9 emoji, B6 invented deadline, A13 inventory ≥ 5, B20 prices.
4. Prompt. If it is a generation habit (logo misspelled), put the fix into the fixed prompt blocks (lock_extra in brand.json) so every ad carries it.
5. Redo spec. The redo spec quotes the Loom in its why field ("Loom 3: try 1 (capsules) was the right idea…") so the review page shows the reviewer you heard him.

4b. Writing the rule (template)

### C19 — <imperative title> — added from <Loom id> <date>
**Rule:** <what to do, brand-agnostic, with placeholders {CORE_PROBLEM} {PRODUCT} …>
**Evidence:** <ad> <try> "<his words>" (<Loom id> [mm:ss]) → <what fixed it> → "<his verdict>" (<next Loom>).
**Enforced:** lint | prompt | QC | process

• One rule = one behaviour. If the new lesson refines an existing rule, extend that rule's text and add a history line instead of a new ID (C2 freeze-dried → fresh).
• Next free IDs: after the highest ID in both rulebooks (3-improve/rules.md here and the wildhorn-static-ads skill's /root/claude-vault/skills/wildhorn-static-ads/wildhorn/lessons-wildhorn.md).
• Add the rule to the index table at the end of rules.md and, if linted, to the lint coverage table.

4c. Extending the linter (hypothetical examples — not rules)

Brand-specific → brand.json (no code): a new ban "banned": {"\\bketo\\b": "avatar already failed keto — don't sell it"}; a new visual check {"if": "\\bpowder\\b", "msg": "A30 we sell capsules, never powder"}. Generic → scripts/lint.py, inside lint(), with the rule ID first in the message (this F7 check is a candidate listed in rules.md, not yet in lint):

        if s.get("template_has_no_product") and not re.search(r"\bNO (product|jar)\b", visual, re.I):
            errs.append(f"#{n}: F7 template has no product — write 'NO product' explicitly in visual")

Then: python3 -m py_compile scripts/lint.py, run lint on the current round's spec file AND on one older spec file to see what the new check would have caught (older files are expected to fail newer rules — that is the point), and record the check in rules.md "Lint coverage".

5. Update the records

• validated.json — every verdict (latest verdict wins, A26). Keep a timestamped backup before rewriting (validated.json.bak-HHMM).
• sell_override.json — {num: try_index} whenever he picks a try that is not the latest (S9→1, S35→2, S42→1, S58→1, S79→1). Frozen forever (A20).
• Protocols log — append the round's two tables ("what worked and why" / "every miss, root cause, permanent fix") on top of the protocols page, newest first.
• Rules page — rebuild (build_lessons.py).
• Stats — stats.py after every round; report first-try %, after-one-fix %, hit rate, generations per validated ad.

6. Regenerate (D11–D12)

• Redo every ❌ ad changing only what he named (A20/A22), passing a validated try as tpl_abs when he liked its layout.
• Apply every dictated fix to every other ad where it fits (D11), against the full rulebook (D11).
• Add the next 10 new ads (D12). Lint → generate → full-res QC → review page. Then wait for the next Loom.

7. Report to the reviewer

A short table per round: ad → verdict → what changed, the new rules (IDs), the stats line, the review link. No victory lap.

---

Case studies — every round, in order

Run 1 (Wildhorn, 79 templates, 9 batches): 0 % first try, 30.3 % of generations validated, 2.5 rounds/ad, 3.3 generations/final ad

L1 — batch 01 written feedback (Adrien, no transcript). Misses: features instead of benefits (#3), added bottom line (#4), too much text (#5), text squeezed the bar (#6), "men over 40" instead of the problem (#7), unclear routine (#8), vague loser lines (#9), "conversion" (#11), changed text although same angle (#14), their headline structure better (#16 verb + body part + deadline), aggressive line kept word for word wins (#17 "Drains beer belly in 14 days"), "X? Try us" not adapted (#18), "your 12 supplements" nobody gets it (#25), customer problem turned into a product problem (#29), sexual concept made non-sexual (#57), great template badly adapted (#63), too many "supplements failed you" overall. → B1 B2 B3 B4 B6 B10 A6 B18 B19 B17 A4 A1 A33.

L2 — batch 02. Wins: #9 Wojak vs Chad "exceptional" (E8); #13 much better (C12); #14 #18 #19 #20 well adapted (A6, B5, C1). Misses: #10 lost the Mars Men illustrated style (C5);

11 joke less logical for the belly (A27); #16 jar crushed (C6). Process: show the first generation next to the new one (D3), feedback loop on good and bad (D4), generate the next 10 (D12).

L3 — batch 03. Wins: #6 changed the scene to show the problem, kept the format (C1); #21 same words, same gag, two nouns (E9); #22 concede the pleasure (B11) — "I would not have put a product" (C7);

24 enemy = mechanism (A28); #25 one word swapped, "apply it everywhere" (E10, D11); #27 #28 #29 better. Misses: #2 changed the side, ugly armless body, logo wrong (C13, A20, D8);

8 man-boobs angle not hit (B18); #23 not readable as a beer belly, text off, competitor logo replaced by a product photo (C1, C7, A20); #26 hair on a non-hair product (A2).

Process: show all versions B1→B4 (D3); next version with all the data since the beginning (D11).

L4 — batch 04. Wins: #8 "amazing — deep into the avatar" (B18, E12); #23 text "on point" (A28); #26 perfect (A2 applied); #29 #30 #31 #32 #34 great (B19, B3, B11, A2). Misses: #2 drew an arm while we sell belly/man boobs (A2 body part); #16 forgot the little orange light on the result (C4, C18); #23 should have replaced the competitor's "Lab" slot (C7);

35 #36 #37 #39 imported their product world — jerky look, brisket, milk pour, chocolate (A2, C2); #37 "we killed the bison… looks like bison shit" (C2, C14, A29); #38 vague rival + vague problem (B5); #40 generic capsule (C2).

L5 — batch 05. Wins: #2 #16 #23 fixed (A2, C4, C7); #36 capsules flanking the jar "really smart" (E11); #37 source → jar with our own mechanism (A29, E11); #40 organs inside the capsule (E3);

41 #42 #46 #49 perfect — their habit → ours, offer slots → our real offer (A2, B20). Misses: #35 format doesn't say "bison organs" in one second (C15); #38 bison hard to recognize (C16);

39 organs not recognizable (C2); #43/#44 "one notch tighter" not meaningful (B9); #47 asked for a radically different version (C8).

L6 — batch 06. Wins: #39 "exactly what I was thinking" — understand the difference maker (E11); #44 B6 "even better than the example" (B9); #47 "the most aggressive — the greatest" (B10);

52 #53 #55 #56 great (B5); #57 hard ad adapted well. Misses: #35 clearer but lost the winning principle (C3); #51 mentioned Autoship (B7); #54 compression shirt (B8); #58 same angle → the bottle should get fat (A6, C6).

L7 — batch 07. Wins: #35 "really understood the game" (C3); #51 #64 #66 #67 great; #70 organs in the capsule "exactly what I imagined" (E3). Misses: #57 not sexual enough / not clearly weight loss (A4, B14);

61 missed the vulva innuendo (A4, A7); #62 watered down to pass GPT — "go for another model" (D1, B10); #63 changed Step 2 (A6); #65 "shirt off since day 21" unreadable + invented timeline (B6);

68 "belly" not "beer belly" (B2); #69 visual didn't match the sexual line (A4); #70 shower copied for no reason (A3). Process: add it to the list of things to solve (D4). → LESSONS.md + blocking lint created here.

L8 — batch 08. Wins: #62 #63 #72 #78 great/amazing (E1, E2). Misses: #57 product not in her hand + text added (A8, B13); #61/#69 kept their innuendo object (A7); #65 invented deadline, flat font, jar off the lounger (B6, A10, A8);

70/#71/#76 organs look like meat; foot not on the scale (C2, A20); #73 text not centred on the eggplant (C9); #74 not on the beer-belly angle (A24); #75 kept the kids' beds (A12);

77 two different bodies (A9); #79 young man, jar not in front of the face, no name on a review format (A11, A8, B12). Process: "you should have a notepad" (D4).

L9 — batch 09 (last of run 1). Wins: #57 #61 #71 #76 #79 validated; #72 B9 amazing (E4). Misses: #65 less good than the example — 3D product and text texture (C10); #69 still doesn't make sense (A7);

70 organs a bit disgusting, no veins (C2); #73 small text too close to the headline, failed again (C9); #77 navel not on the cut (A9); #57 product a bit small (C17).

Key instruction: "#75 — you could have done this from the first try… every pixel of the scene" → A13 pixel inventory, mandatory and linted.

Sam batch (83 new templates, 10 at a time, rulebook + lint from the start): 71 % first try, 87 % after one fix, 100 % final, 62 % of generations validated, 1.4 rounds/ad, 1.6 generations/validated ad

SL1 — round 1 (S1–S10): 6/10 first try. Wins: S1 bananas → shrinking belts (flip their object); S3 puzzle jar with organs + one job per organ; S4 text-only by script (zero text errors); S5/S6 hidden-hormone checklist mapped onto our mechanism; S10 same buyer → their words (E5, A6). Misses: S2 GLP-1 faces missed the belly (A14); S7 button popping tamer than worms (A15); S8 silhouettes faced away (A16; SL2: before image more aggressive, C18); S9 "You can't farm this" unclear to a stranger (B14). Question: S10 "why sexual and not beer belly?" → A6 then, A14/A31 now.

SL2 — round 2 (S2–S40): 26/34 new ads first try. Wins: S12 S16 S20 S21 S22 S27 S32 S36–S38 (E7); S23 moon (E6); S7 S9 now validated (A15, B14); S28 arcade fusion (E13); S32 "so aggressive" (B10). Misses: S11 jar not on her head — fallback dropped a placement (D6, A21); S14 flipped a line that already worked (A19); S15 shoelaces (A18); S17 brown organs (C2 rewritten); S18 capsules next to the beer instead of the pour (A17); S24 "felt the difference" too soft (B16); S29 words too complicated (B15); S33 bed copied + phone badly held (A3, C11); S35 "Outperform your 30s" (A14, A24).

SL3 — round 3 (redos + S41–S83): 31/43 new ads first try; QC fixed 13 before he saw them (logo misspelled ×4 → D8; template leftovers ×2 → C5/F11; Seedream full ads ×4 → D7; emoji box → D9). Wins: S15 S17 S24 S33 fixed by the round-2 rules (A18, C2, B16, A3+C11); S21 S44–S49 S51 S53 S54 S59–S61 S63 S71–S73 first pass. Misses: S11 the woman's FACE is the product (A21); S18 S58 S79 regenerated tries he had validated (A20); S18 powder makes no sense for capsules (A30); S29 style changed when only wording was wrong (A22 — root cause found later: the spec pointed at S31's template, D13); S74 two lean bellies (A23); S50 sexual fruit (A23); S42 S52 S56 S80 off-angle or vague (A24); S57 missed the woman angle (A24); S55 she initiates → she should be on top (A32); S79 man boobs, not belly (A31); S42 redo "OK" with no image (D10).

SL4 — round 4 (the 12 left). Wins: S14 S18 S29 S50 S52 S56 S57 S74 S80 — the round-3 rules fixed them first time (A19 A30 A22 A23 A24); S62 watercolor copied as is (E7); S42 try 1 validated (A26). Miss: S11 for the 3rd time — kept regenerating a scene; "review the ad by screenshot and understand the model they use" → A25 (edit the template minimally, masked product swap, text by script, keep 9:16).

SL5 — S11 only (no transcript). "Great job on the last one" — A25 worked. 83/83 validated.

---

Loom coverage matrix

Every distinct piece of feedback in every transcript (feedback/loom-b2…b9.txt, loom-s1…s4.txt) plus the two reviews without a transcript (L1 = batch-01 written feedback on /protocols; SL5 = Sam Loom 5 verdict on /protocols) → the rule that encodes it. Type: W win (repeat) · M miss (rule) · P process instruction · V plain verdict (validation only, no lesson). Rules marked † were added from the Loom audit (they existed only as page notes or not at all in LESSONS.md).

Coverage: 231 feedback items (124 misses · 73 wins · 20 process instructions · 14 plain verdicts) · 0 without a rule. The 14 plain verdicts (V) carry no lesson; they only update validated.json. Rules with † were written from the Loom audit. 24 of them are shared with the Wildhorn skill (same IDs as /root/claude-vault/skills/wildhorn-static-ads/wildhorn/lessons-wildhorn.md): A27–A31 · B17–B19 · C12–C18 · D11–D13 · E8–E13. Three exist only in this skill: A32 (staging acts out the line), A33 (favor aggressive angles), B20 (real facts only). F1–F13 consolidate the QC logs (not Loom feedback); R1–R14 restate the blueprint system rules.

When a new Loom arrives, append its rows here (same columns). A row with an empty Rule column is a failure of the loop.

Transcription garble dictionary (Wildhorn Looms — check every Loom for the same kind)

Rule: never act on a garbled word without checking the ad on screen at that timestamp; when unsure, ask once in the round report.