AINA Academy · Frontend & Product-Experience Lane · 2026-08-06

Every screen, walked as a real learner — what's rough, and the plan to polish it

Evidence-first inventory of the whole surface — marketing landing through lesson, practice, progress, checkout — with root causes verified in source, three sequencing options, and one recommendation. Nothing here has touched code yet: this is the plan for your ruling.

Signed-in evidence: staging under real auth, today's green S5-r1 run (31116101411), head d869e4cd Everything else: local walk of the same code, desktop 1440 + mobile 390 Staging untouched (S5 campaign in flight)

Status: PROPOSED. Plan of record: docs/plans/frontend-polish-lane.md. This lane keeps ONE plan — new asks append, never fork.

The one-paragraph verdict

The learner loop is real and works end-to-end — the lesson shell is genuinely strong. But the product currently lies to learners in small ways (invented dimension scores, a decorative readiness number that reads as yours), leaks internal machinery (raw enums, config strings, snake_case keys as learner copy), has invisible text on its three dark info cards, and dead-ends anonymous visitors on four routes. The marketing face is still the pre-V6 page you already ruled superseded, with no footer, no privacy link, and analytics that start before consent. All of it is fixable in small, reversible, screenshot-reviewed PRs — and the honest fixes should land before the new face does.

Surfaces inventory

Surface	Route	Condition	What the walk showed
Marketing landing	`/`	superseded	Old MarketingLanding, copy variants rotate per visitor (career-switchers / team-leads). No footer, no privacy/terms, no cookie consent; PostHog inits for every visitor. V6 (PR #222) is the ruled replacement.
Sign-in / Sign-up	`/sign-in` `/sign-up`	rough	Anonymous visitor gets the full signed-in app chrome + 6-item nav; no fallback when the auth widget can't load.
Home dashboard	`/` signed in	rough	Works; "Welcome back, learner" placeholder, account-deletion card on the home surface, empty em-dash tiles, stale readiness number.
Diagnose → snapshot	`/diagnose`	dishonest	Flow is good; but per-dimension scores are fabricated from the overall score, and the hero's fake "68/100" reads as yours.
My Path	`/path`	fine	Honest gate; loading state is a bare sentence.
Lesson shell	`/lesson`	strong, cluttered	Best screen in the product. Same takeaway rendered 3–4× at once; "Lesson 00"; internal chips visible on mobile.
Practice arena	`/practice`	leaks	"Mode: Pure Simulation · Skin: Instructor Studio" shown as a chat message; scenario card text near-invisible.
Progress	`/progress`	broken card	Raw enum "NEXT_IN_SEQUENCE" in the heading; next-best-action card text invisible; "+0pp" jargon.
Settings (DAT-170)	`/settings`	anon dead end	Signed-in consent card shipped well; anonymous visitors get a title and then a blank page.
Checkout	`/checkout`	bare	"AI FLUENT PLUS · $349" — different product name than the landing; no what's-included, no trust signals, no way back.
Library	`/library`	fine	Clean, honest empty state.
Playground	`/playground`	anon dead end	"Loading missions…" forever for anonymous visitors — no redirect, no error.
Memory inspector	`/memory`	anon dead end	Lesson chrome + "SCORE 0/100" for anonymous; loads forever.
Workspace	`/workspace`	fine	Redirects home.
Lesson preview	`/lesson-preview`	broken card	Slide subtitle invisible on the dark card; the caption repeats the invisible line just below it.
Mobile / Admin apps	—	not walked	Expo app and admin need their own pass; excluded from this proof set.

Cross-cutting on mobile: the bottom nav's six items wrap "Progress" onto a lonely second row on every page, and the header logo sits in a white box on the cream header.

P0 — dishonest or broken for real users (all root-caused in source; smallest, highest-impact fixes)

P0-1

Dark info cards render near-invisible text — three surfaces, one root cause

Progress "Next best action", Practice scenario, and Lesson-preview slide all show barely-visible text on the navy card. Source says cream-on-navy — the real render says otherwise, so the failure is at the token/color-mix layer. One fix at the token level, not three per-card patches.

Progress page: next best action card with near-invisible text and raw enum NEXT_IN_SEQUENCE — **Progress, staging, real auth.** The dark card's heading and body are unreadable — and the heading that IS legible says "NEXT_IN_SEQUENCE" (P0-3).

Lesson preview slide with invisible dark subtitle on dark card — **Lesson preview.** The slide subtitle is invisible on the dark card; the caption below repeats the same sentence.

P0-2

The snapshot's per-dimension scores are invented

dimensionScores() fabricates "Tool fluency = score+8 · Output review = score−10 …" from the single overall score and presents them as measurement — at score 0 every bar clamps to "10%". Real dimension evidence exists in the progress API. Either wire the real data or label the panel honestly; a learning product cannot fake precision to learners.

Assessment snapshot: fluency by dimension bars all reading 10%, snake_case focus areas — **Snapshot, staging, real auth.** All four "dimensions" read 10% (the clamp floor of invented numbers). Also visible: focus areas as raw keys — `tool_judgment`, `workflow_design`, `verification_habits` — and "Mesa assessment features" (internal system name) in learner-facing rationale.

P0-3

Internal machinery leaks into learner copy

Raw enum NEXT_IN_SEQUENCE on Progress; snake_case focus areas on the snapshot; "Mesa" jargon; practice chat opens with "Mode: Pure Simulation · Skin: Instructor Studio"; mobile lesson shows demo:prompt_task_success_criteria and lesson_id / learner_work chips. One label-map + copy pass, one PR.

Practice arena mobile: internal mode/skin config shown as chat message, dark scenario card barely readable — **Practice, mobile.** Internal config surfaces as the first "message" a learner reads; the scenario card repeats the contrast bug.

Lesson mobile: raw demo tokens and lesson_id/learner_work chips visible — **Lesson, mobile.** `demo:prompt_task_success_criteria` and raw chips visible below the mission card.

P0-4

Anonymous dead ends on four routes

/settings renders a heading then nothing; /playground and /memory spin forever; sign-in renders the app shell with no fallback. Every guarded route should either redirect to sign-in or say plainly what's needed. One route-guard consistency pass.

Settings page blank for anonymous visitor — **/settings, anonymous.** "Manage how AINA works for your account." — then a blank page.

Playground stuck loading missions forever — **/playground, anonymous.** "Loading missions…" never resolves.

P1 — coherence & trust (your taste calls live here)

P1-5

V6 landing reconcile — the centerpiece

PR #222 carries your ruled V6 direction (pi.dev-style sticky product-demo scroll) and is now behind main. Plan: rebase onto current main — reconcile, never fork — wire its CTAs to the real /diagnose and /dashboard routes (in-page anchors today), design the footer + DAT-171 consent banner with the page, retire MarketingLanding in the same motion, keep the E2E harness selectors true in the same PR. I'll bring 2–3 cheap in-context variants of the contested calls (hero copy angle, CTA strategy) for you to react to first.

Current pre-V6 marketing landing, team-leads copy variant — **Current landing (team-leads variant), local walk.** Nine-line serif hero, abstract product card, no footer, no privacy link, no consent — and a second copy variant (career-switchers) rotates in for other visitors. This whole page is scheduled to be replaced by V6.

P1-6

The funnel tells three different stories

Landing sells "$349 · L1-L5 Spine + Credential" via waitlist; checkout sells "AI FLUENT PLUS · $349"; the header offers Sign in; on scroll, two identical waitlist CTAs stack on top of each other. One product name, one primary CTA, one anon → sign-up → paid story. Blocked only on your naming/positioning call (Question 3 below).

Checkout page: bare AI Fluent Plus card at $349 — **/checkout.** A different product name than the landing, no what's-included, no refund/security reassurance, no back path.

P1-7

Home dashboard truthfulness

"Welcome back, learner" (name is available); account deletion parked on the home surface (belongs in /settings — natural DAT-170 extension); readiness circle stays 0 after a lesson moved fluency to 3% on Progress; em-dash empty tiles.

Signed-in home dashboard with placeholder greeting and delete-account card — **Home, staging, real auth.** Placeholder greeting up top; "Delete my account & data" as a first-class home card at the bottom.

Home dashboard mobile: tall empty stat tiles and delete card — **Home, mobile.** Full-width empty tiles stack into a long scroll that ends at the deletion card.

P1-8

Diagnose hero shows a fake score as if it were yours

"YOUR READINESS · 68/100 · Building fluency · Confidence 56-78%" is decoration rendered before you've answered anything. Label it as a sample report or replace it with product truth.

Diagnose hero with fake 68/100 readiness panel — **/diagnose, staging.** The panel reads as the learner's own score. It's a hard-coded prop.

P1-9

Lesson screen repeats itself

The same takeaway paragraph renders up to four times on one screen (chat body, right-rail KEY TAKEAWAY, MISSION CARD, "The durable move"). And lessons are zero-indexed learner-facing ("Lesson 00"). Collapse the duplication; human-number the lessons.

Lesson shell with the same takeaway text repeated in four places — **Lesson, staging, real auth.** Strong bones — the duplication is the tell that three panels are fed the same string.

P2 — polish

P2-10

Mobile nav + brand details

Six bottom-nav items wrap "Progress" onto its own row on every mobile page; the header logo ships with a white background on the cream header; page-title styles mix "—" and "-"; path/playground/memory load as bare sentences instead of skeletons; checkout needs its what's-included + trust block.

Sequencing — three options, one recommendation

Recommended

Option A — truth first, then the new face, then coherence

Wave 1 (P0): 2–3 small PRs — contrast root-cause · honest dimensions · label/leak pass + route guards. Each with screenshots for your label, sequenced between S5 rounds. Wave 2 (V6): #222 rebased and wired, footer + CMP banner designed in, old landing retired; you react to 2–3 in-context variants before the PR. Wave 3 (P1/P2): funnel naming, home truthfulness, lesson dedup, then polish. Why: the P0s are honesty bugs live in front of real users today, they're already root-caused and tiny, and V6 deserves your unhurried taste pass — not a rush job.

Option B — V6 first

Biggest visible change lands first; P0 fixes follow. Cost: learners keep hitting invisible text and invented numbers meanwhile, and the #222 rebase gets riskier if Wave-1 style fixes land after it anyway.

Option C — one big polish PR

Rejected: un-reviewable under the founder gate, one giant QA receipt, and it collides with the staging campaign's merge sequencing.

Three questions for you (nothing blocks Wave 1)

Launch audience of record — the old landing rotates career-switcher and team-lead pitches; the meta description says working professionals. Which voice does V6 speak in? (blocks Wave-2 copy)
CMP pick — you said you have consent-management tools chosen (DAT-171). Which one(s), so the V6 footer/banner integrates it rather than hand-rolling. (blocks the consent part of Wave 2)
Product name + primary CTA — "AI Fluent Plus" vs "L1-L5 Spine + Credential"; waitlist vs direct sign-up on the landing. (blocks the Wave-3 funnel PR)

Rails this lane honors, every wave

All product frontend is founder-gated (D-039 am.2/3): you personally see screenshots before anything merges; I never self-apply your label. The E2E journey harness navigates the real UI — aria-labels, routes, roles, CTA text are load-bearing proof selectors; any rename updates scripts/e2e/headless-full-journey.ts in the same PR, impact noted in the PR body. Merges to main redeploy staging — sequenced with the S5 campaign session between rounds; deploy.yml is never dispatched. Design language extends global.css + @aina-academy/ui tokens — no forks. One plan (docs/plans/frontend-polish-lane.md); asks append to it.