Design review · Phase C · 2026-06-11

Kashi — Deck visual + brand audit across 3 decks.

Visual hierarchy, AI slop patterns, brand consistency, print-PDF readiness. The biggest finding: EN decks and JA deck use DIFFERENT brand palettes — evergreen #324532 (EN) vs emerald-700 #047857 (JA). Same product, different visual identities.

Provenance: Phase C = Claude-internal design review across all 3 deck HTML files. The gstack /design-review skill is built for live-browser QA (visual inconsistency / spacing / hierarchy / AI slop / slow interactions) — it requires a running app or browser session, not static slide HTMLs. This audit is the static-analysis substitute: reading each deck's CSS + slide structure + content density and flagging visual / hierarchy / slop issues per slide. Findings are anchored to specific file:line pairs.

01 / Master visual scoreboard3 decks × 5 design dimensions

Deck Visual hierarchy AI slop Brand consistency Print-PDF ready 4-dimension visibility
deck-en-alchemist-interview 6/10 5/10 8/10 8/10 3/10
deck-en-10slide 7/10 7/10 8/10 9/10 3/10
kashi_pitch_v3_final (JA) 8/10 8/10 5/10 9/10 5/10

Read this scoreboard like: the JA deck is the best-designed of the 3 (strong hierarchy, real mock visuals, low slop) — but uses a different brand palette than the EN decks. The two EN decks are visually clean but neither dedicates a slide to any of the 4 dimensions Justine named (problem, success, moat, why-not-GPT). All 3 decks fail "4-dimension visibility" — none let a partner reading the deck cold answer the 4 questions in <60s each.

02 / Brand palette inconsistencyThe EN ↔ JA visual identity gap

ElementEN decks (alchemist + 10-slide)JA deck (v3_final)Canonical
Primary green#324532 evergreen (deep, muted, warm)#047857 emerald-700 (saturated, cooler)Memory says emerald-700 is canonical for typographical accents; evergreen is the "deep brand". Both legitimate but inconsistent across surfaces.
Heading fontInter 700Fraunces (variable) + Zen Kaku Gothic New 700JA uses canonical Fraunces wordmark (typography-spec-v1 §4.1); EN drops Fraunces for Inter (faster, more enterprise-modern). Trade-off Justine made consciously.
Body fontInterZen Kaku Gothic New + IBM Plex SansJustified — JA needs CJK-optimized rendering.
BackgroundCream #F5F0E6 + Evergreen dark slidesLight gray #E2E5E9 page + white slidesDifferent metaphor. Cream = warm B2B SaaS. White-on-gray = traditional pitch deck. EN feels more modern.
AccentCoral #E8836BEmerald-100/50 (light green)Coral is more attention-grabbing; emerald is more in-family. Different vibes.

Recommendation: For v5 — unify on the EN palette (Inter + evergreen + cream + coral) across BOTH EN and JA. JA can keep Zen Kaku for body but use Inter for kicker/labels and evergreen for headers. This costs visual polish on the JA deck (loses Fraunces wordmark warmth) but gains brand coherence for partners who see both. The Alchemist + J-StarX partners are global; they'll see both eventually. Alt: keep two palettes, document the rationale as "EN = modern enterprise / JA = warm traditional" — but this adds confusion when a partner asks "which is the real Kashi?"

03 / Deck-by-deck design findings

deck-en-alchemist-interview.html (Jun 4, 12 slides, talking-head format)

Slide / lineFindingSeverity
S6 / L362-369AI slop: 2 cards "What we do / What we don't" — generic structure. The "What we don't" card just lists 3 negatives ("Judge people. Predict who will quit. Diagnose emotions.") in 14pt. No visual weight, no hierarchy. Reads as defensive disclosure, not as moat.HIGH
S7 / L375-385Good design: compare layout (left = existing tools, right = Kashi) is a SOLID hierarchy. Dark Kashi card vs light competitor card creates instant contrast. Keep this pattern.POSITIVE
S8 / L393-404AI slop: 3 cards "No emotion recognition / No employee scoring / Content never stored" — list of negatives. Same problem as S6. A "Why we built it this way" angle would be stronger.MED
S10 / L433-444Visual issue: founder bullets (3 of them) + pull quote — too dense for 1 slide. Pull quote ("People do not leave suddenly. They become silent first.") deserves its own slide.MED
S11 / L450-461Good design: 4-card traction grid with big numbers (v1 / 908 / 25% / →) is visually strong. Numbers carry the load.POSITIVE
Cover S1 / L292-305Brand strength: Wordmark + tagline + identity meta in a clean dark slide. Cover is solid.POSITIVE
Cross-cutting4-dimension visibility FAIL: No dedicated Moat slide (S6 = problem reframe, not architecture moat). No dedicated Success slide (S11 = traction, not 3-phase). No dedicated Why-not-GPT slide.HIGH

deck-en-10slide.html (May 31, 10 slides, J-StarX submission format)

Slide / lineFindingSeverity
S5 Why now / L127-138Good design: 3-card Loss / Obligation / Regulation grid is high-impact. Visual hierarchy correct (kicker → headline → cards → footer). KEEP this pattern.POSITIVE
S6 Moat / L140-152Visual hierarchy issue: Slide title "Our moat isn't a feature. It's how the product is built." is good. But the 4 bullets that follow describe properties, not the architecture moat V3 prescribes. The footer "Hard to copy without redesigning the entire product" is in a different style (small italic). Eye lands on bullets, not on the structural argument.HIGH
S7 Traction / L154-181Density issue: 3 traction bullets ABOVE + 3-card 6-month plan BELOW = 1 slide doing 2 jobs. Either split into "Traction" + "6-month plan" or compress.MED
S8 Market / L183-209Good design: 3 stat columns (TAM/SAM/SOM) with big numbers + pricing ladder. Strong visual hierarchy. Keep.POSITIVE
S9 Team / L213-219Visual + content issue: Founder slide has 4 lines of dense bio + LinkedIn URL placeholder ("[paste LinkedIn URL here — see BRIEFING note]"). Placeholder MUST NOT SHIP. Also: bio reads as paragraph not as scannable. Convert to bullets or a card grid.HIGH
S10 Ask / L221-233Strong closer: 3 asks in a list + "Japan first. Built to travel." + tagline closer. Good shape.POSITIVE
Cross-cutting4-dimension visibility FAIL: Same as Alchemist deck — no dedicated Moat / Success / Why-not-GPT slide.HIGH
Print-PDF checkGood: @media print rules set @page size:landscape margin:0, each slide page-break-after:always. Print-PDF works cleanly via ⌘P.POSITIVE

kashi_pitch_v3_final.html (May 27 JA, 23 slides)

Slide / lineFindingSeverity
S06 Product / L580-694Strong design: 3 mock SVG visualizations (Exec Cross-team view / Manager Mirror / Member View) are the best slide in any of the 3 decks. Concrete product evidence. KEEP.POSITIVE
S07 Boundary / L699-715Strong design: 3-column table 見えるもの / 見せないもの / なぜ — clean architecture-not-policy framing. Strong content. Visual is also clean.POSITIVE
S11 Market / L793-817Strong design: 3-stat row (SOM / SAM / TAM) with descriptive footnotes. Numbers carry the load. Matches EN 10-slide S8.POSITIVE
S13 Founder / L920-940Visual + content issue: Title "単独でここまで作った" is visually emphasized — which makes the V3-forbidden "solo" framing more prominent than it should be. The same slide with V3-aligned "共同創業者級のチームメイトと共に" headline would land very differently.HIGH
S14 Architecture / L940-1000Good content, dense visual: Detector + privacy + access boundaries explained. Heavy with bullets. Could benefit from card-based layout matching S07 boundary slide.MED
S16 Regulatory / L1030-1060Visual issue: Article 5 EU AI Act + APPI bullets in a card. The "勝負" framing makes it read as a competitive angle vs partners — V3 reframes as compliance posture.HIGH
Cross-cutting4-dimension visibility PARTIAL: Problem slide (S02 voices-too-late) present and good. Success (S08 + S09 + S11) partial — 6-month + market sizing exist but no V3 Phase 1/2/3 structure. Moat (distributed across S05 + S07 + S14) — no single moat slide. Why-not-GPT — not present.HIGH
Page numbering "/ 23"Design issue: 23 slides for an interview deck is too many. McKinsey-style "/ 23" page numbering reinforces the wall-of-slides feeling. Consider splitting into "main 16 + appendix 7" with appendix marked clearly so interviewers know they can skip.LOW
Brand inconsistencyUses emerald-700 #047857 + Fraunces + Zen Kaku — does not match the EN decks' evergreen + Inter. Visible to any partner who sees both decks.HIGH (cross-deck)

04 / AI-slop patterns to drop in v5

Specific generic-deck patterns visible across the 3 decks. These read as "another B2B SaaS pitch" rather than as Kashi-specific. v5 should avoid:

05 / V5 deck structure recommendation

Apply the codex-audit + gemini-audit findings + the design observations above. Recommended v5 slide-count + slide structure (target = same across all 3 decks for brand coherence):

#SlideWhy this slideV3 anchor
1CoverTagline + identity meta
2Hook — 退職代行 / Resignation agencies as visible symptomV3 mandates this hook explicitlyV3 L23-25, L46-47
3Problem — more communication, less visibility (4 cards of early signs)Concretize the dialogue breakdownV3 L46-47, L96-101
4Solution — Kashi reads the SHAPE, not content (3-step pipeline)Core mechanism explainerV3 L50, L106
5Product — 3 mock visuals (Exec / Manager / Member views)Concrete product evidence (JA deck S06 pattern)
6Boundary — what we see / don't see / whyArchitecture-not-policy boundaryV3 L106, L118
7NEW — Moat (3 architectural reasons)verbatim V3 3-beat as 3 cards: auditable+reproducible+content-blind / persistent state / doctrine in data layerV3 L144-146
8NEW — Why won't GPT eat usReframes the same 3-beat as commodity-LLM-defense (Microsoft trap addressed honestly per codex finding)V3 L52-58, L108-114
9Why now — 3 forces (Loss / Obligation / Regulation)EU AI Act Annex III §4 + MHLW + 退職代行 as visibility-cost signalV3 L60, L116
10Market — Tokyo SMB 50-500 ICP + TAM/SAM/SOM + reseller channelSpecific ICP per V3V3 L62-63, L120, L154
11NEW — Success (3 phases with concrete numbers)Phase 1 (12mo: 5 pilots, SLA 80%+, NPS 7+, 1 renewal) / Phase 2 (24-36mo: 50 customers, 30 via reseller) / Phase 3 (5yr: default trusted lane)V3 L152-154
12Traction — 908 / 25% / pilot conversations + honest pre-pilotTime-anchored proof per V3V3 L66, L126, L162
13Founder fit — Justine + co-founder-level teammateV3-aligned solo→co-founder framing; lived-problem credentialV3 L30, L48, L104, L158
14Ask / Closer"Before silence becomes resignation."V3 L32, L128

14 slides for EN long (Alchemist + investor deck), 12 slides for EN short (J-StarX submission — collapse Moat + Why-not-GPT into one), 16 slides for JA (add Appendix-marked 2 slides for market math + 5 detector specs).

06 / Print-PDF readiness

All 3 decks have @media print rules with @page size:landscape and page-break-after:always per slide. Tested mental walkthrough: