AdPilot · v4 design system · decision + full-fidelity mocks

Calm Studio — chosen direction, built out

You picked B · Calm Studio with three tweaks, all applied below: a dark theme variant, borrowed components where Calm needs density (Sentry-style tables, telemetry numerals, ⌘K), and more data on screen than pure-Apple minimalism. All six surfaces are mocked at full fidelity.

✓ Decision recorded
Calm Studio v4: Apple-calm light as the default, a full dark variant, and dense data components borrowed from the other directions wherever numbers need room. This page is the reference for the T-029 implementation card.

Design system

tokens · light + dark · borrowed components

Base is Apple-calm: soft-shadow white cards on warm gray, SF-style type, one blue action. Where the product needs density, three borrowed pieces plug in: Sentry-style dense tables (campaigns, runs, audit), telemetry numerals (monospaced tabular KPIs, borrowed from the Starlink direction), and ⌘K keyboard-first chrome (from Studio Mono).

Light theme (default)

#F5F5F3 paper · #FFFFFF card · #0071E3 action · #1D7A53 win · #9A6B1B warn · #C0392B risk · hairline #ECECEA · radius 14 · soft shadow

Dark variant

#0E0E11 paper · #1B1B20 card · #0A84FF action · #30D158 win · #FFD60A warn · #FF453A risk · hairline #2A2A30 · no shadows, hairlines only
Type: SF system stack; numerals in monospaced tabular for anything money (the one Starlink borrowing used everywhere). Density rule: cards stay calm; tables go dense (12.5px rows, mono figures, status chips) — the Sentry borrowing. Every screen is reachable by ⌘K — the Studio Mono borrowing.

1 · Dashboard

calm shell · dense table · telemetry KPIs

Apple-calm greeting and KPI hero up top (with monospaced numerals), then a dense sortable campaigns table below — the "more info" you asked for lives in the table, not in chrome.

Good morning, Navid

Week 37 · your ads returned $4,218 across 3 campaigns.

Returned
$4,218▲ 12.4%
Spent
$1,09272% of plan
ROAS
3.86▼ 0.2
Orders
214▲ 31

Campaigns · last 7 days

CampaignPlatformSpendReturnedROASCPAStatus
Summer DropMeta$380.11$1,790.204.71$4.10RUNNING
Summer DropTikTok$299.02$1,166.183.90$5.60RUNNING
Ramen Kit LaunchTikTok$412.90$545.031.32$18.40WATCH
Ramen Kit LaunchMeta$0.00———DRAFT
Ramen Kit Launch · TikTok — ROAS under 1.5 for 3 days. Copilot proposes shifting 20% of its daily budget to Meta (guardrail: ±30%).
Proposal #48 · waiting for your approval

2 · Studio

one prompt box · mic · advanced panel open

Beginners see one input. The Advanced panel expands for aspect presets, model, voice, hooks and variants — nitty-gritty lives here, hidden by default. Mic sits in the field: speak instead of typing.

Create

Summer Drop · wave 2 — describe it, or press the mic and say it.

Cozy late-night ramen ad — warm kitchen, steam rising, "ready in 3 minutes", order-now end card…🎙
⌘M to talk · transcription lands in the field, editable before generate

Advanced

1:14:5 Meta9:16 TikTok16:9
3 variants
Flux Pro ▾
Warm · Female ▾
Question hook ▾
text-heavy, green tint
Brand kit Ramen Bar · auto-applied Follow brand kit ✓

3 · Platform previews

how the customer will actually see it

Every creative renders inside its placement: Meta Feed (4:5) with feed chrome, Story/Reels and TikTok (9:16) with side-rail actions, Google Display banner. v1 ships hand-built frames; the Go API can later wrap Meta's /generatepreview HTML for true-platform renders (documented in T-022).

Summer Drop · variant B

Meta FeedStory / ReelsTikTokGoogle Display
Nami Ramen
Sponsored · 🌐
⋯
18-hour broth. 3 minutes of your time. The kit that made takeout optional 🍜 See more
$24 · feeds 2
Order Now
Nami Ramen Kit — feeds 2nami-ramen.com
👍 1.2K84 comments · 26 shares
👍 Like 💬 Comment ↗ Share
META FEED · 4:5
namiramen2h
✕
Send message ❤️
Order now · $24 ↗
Reels placement · same unit on Stories
REELS / STORY · 9:16
FollowingFor YouLive
+➕ ❤️
12.4K
💬
342
🔖
2,180
↗
1.1K
Sponsored
@chefmika Follow
POV: a 3-minute kit that beats your takeout order 🍜 #ramen #foodtok #ad
♬ original sound – chefmika
Shop now
TIKTOK IN-FEED · 9:16
🍜
Nami Ramen · Sponsored
Ramen Kit — Ready in 3 Minutes
Restaurant-quality broth at home. Feeds 2, ships nationwide.
Shop now
GOOGLE DISPLAY · RESPONSIVE

Anatomy per current platform units: TikTok in-feed spec (right rail like/comment/save/share, Sponsored + caption + sound, ads use Commercial Music Library), Meta feed ad unit (Sponsored under page name, CTA + headline + domain, counts above pill actions), Meta Story CTA + message pill, Google responsive display card. In-product, real previews switch to Meta /generatepreview HTML when an account is connected; frames here are the no-account fallback and are pixel-checked against the live apps each release.

4 · Automations

plain-language rules · spend-cap pause · run history

The builder reads like a sentence: When… crosses… then… — created by form, by typed request, or by voice ("pause my ad if it costs more than $100"). Budget-touching actions require confirmation. Run history is a dense table.

Automations

When spend on Ramen Kit Launch · TikTok crosses $100
then pause campaign and notify me
Checks every 30 min · cooldown 6 h · max 2 actions/day · pause executes in sandbox until a live account is connected
When ROAS on any campaign stays below 1.5 for 3 days
then propose budget shift −20% · needs approval (budget action)

Run history

WhenRuleObservedActionResult
Mon 09:30Spend ≥ $100$102.40Pause + notifyPAUSED
Sun 18:00ROAS < 1.5 × 3d1.32Propose −20%AWAITING OK
Sun 18:00Spend ≥ $100$61.10—NO TRIGGER

5 · Library

gallery with a real edit path back to Studio

Every asset carries its context (prompt, seed, aspect, version lineage). Hover actions: Edit in Studio (pre-fills the whole generation), Remix (same prompt, new seed), Duplicate, Copy prompt. The version strip shows parent/child history.

Library

Search prompts, tags, campaigns…
AllSummer DropRamen KitVideoDrafts

Version lineage · Summer Drop (feed)

v1 · Sep 6→v2 · Sep 8→v3 · current regenerations link to their parent; compare any two side-by-side

6 · Copilot

chat ops · voice in · numbered proposals · approvals

The conversational operator: type or speak, get a numbered proposal card with per-platform reasoning, approve or adjust. Voice turns are transcribed into editable bubbles (the wave chip).

Copilot

I have $500 for September — ramen kits, mostly TikTok people. 🎙 0:07
Here's the plan. TikTok carries discovery; Meta retargets the clickers. Daily budgets respect every platform minimum.
Proposal #48 · Ramen Kit · September
TikTok
$300 discovery · 9:16 UGC-style
Meta
$150 retargeting · 4:5 feed
Reserve
$50 shifts weekly by ROAS
KPIs: ROAS ≥ 2.0 · CPA ≤ $8 · I'll rebalance weekly within ±30% and tell you why, every time.
sandbox launch · 0 credits until real account
and pause the tiktok one if it costs more than $100
Done — automation created: "pause Ramen Kit · TikTok when spend ≥ $100" · checks every 30 min, you'll be notified. Find it in Automations.
Message AdPilot…🎙

Dark variant

same system · hairlines instead of shadows

Full dark theme, same layout and components — tuned for data readability, not just looks: secondary text lifted to #BCC0C9-class luminance, table cells near-white at medium weight, high-chroma status chips (mint #3DDC84 / amber #FFD426) on tinted grounds, brighter hairlines (#34343D) separating cards from the near-black canvas. Shown: Dashboard and the Automations run history, where dense data must survive the dark.

Good evening, Navid

Week 37 · $4,218 returned · one campaign needs a look.

Returned
$4,218▲ 12.4%
Spent
$1,09272%
ROAS
3.86▼ 0.2

Automations · recent runs

WhenRuleObservedResult
Mon 09:30Spend ≥ $100 → pause$102.40PAUSED
Sun 18:00ROAS < 1.5 × 3d → propose1.32AWAITING OK
Ramen Kit · TikTok paused by rule "Spend ≥ $100". Resume when ready — I'll re-baseline pacing.

7 · Workflow studio

LLM plans → code executes · execution graph · step logs

The workspace-agent pattern: a GPT layer turns intent into a workflow plan, and a code-execution layer runs each deterministic step — predictable, cheap on tokens, small context per step, small surface area for mistakes. The execution graph shows exactly after which step what is set, and the run log below shows from-which-point-to-which-point everything that happened in history.

Welcome flow · ramen-kit-september

LIVE · v4

Execution graph · after each step, what is set

TRIGGER
Webhook in
POST /hooks/lead · source: landing page
→
GPT STEP
Segment intent
classifies lead: gift · lunch · late-night · output: JSON enum
→
CODE STEP
Build audience
deterministic · maps enum → platform audience payload · validated by schema
→
BRANCH
A / B by intent
gift → variant A · late-night → variant B
→
ACTIONS
3 fanned out
landing variant · ads experiment · sales webhook

LLM proposes, code disposes — the GPT step only emits a schema-validated enum; everything downstream is deterministic and replayable.

Run log as a graph · waterfall timeline

run #841 · x-axis = time from trigger · one lane per step · color = kind, red = failed attempt trigger gpt code branch action
Webhook · trigger
12ms
Segment intent · gpt
820ms · 412 tok · $0.001
Build audience · code
4ms · schema ✓
Branch A/B · branch
1ms → variant A
Landing enroll · action
96ms · arm A
Sales webhook · action ∥
61ms · fan-out lane
0200ms400ms600ms800ms1.0s1.2s

Retry lineage · run #840 — the failed attempt is visible on the timeline, not buried in rows

Segment intent · gpt
✗ schema invalid · 610msretry 1 · 1.1s ✓ · resumed from checkpoint

Last 20 runs

17 ok · 2 retried · 1 failed · hover a bar to jump to its log rows

Run log · detail — same run, row per step (expandable in the product)

RunStepKindIn → OutTookResult
#841Webhook: lead capturedtriggerPOST /hooks/lead → event 9f2…12msOK
#841Segment intentgpt"gift for my brother, spicy" → {intent: "gift"}820msOK
#841Build audiencecode{gift} → audience_payload v3 (schema ✓)4msOK
#841Branch A/Bcode{gift} → variant A1msOK
#841Landing variant · enrollaction→ experiments/ramen-sept · arm A96msOK
#840Segment intentgpt"…hungry at 1am…" → schema ✗ · retried once1.1sRETRY→OK

Every GPT step emits into a schema; every code step is pure and replayable; a failed step re-runs from its checkpoint, never from scratch.

8 · Landing pages + Experiments

a page per audience · winners earn budget

Workflows can generate a different landing page per audience and intent, and the experiment layer learns which kind of person converts on which kind of page — then feeds the winner back into targeting.

Experiments · ramen-kit-september

2 RUNNING

Landing pages · generated per audience

AudiencePage angleCVR
Gift buyers"The gift that gets eaten" · warm, ribbon imagery4.8%arm A
Late-night scrollers"Ready in 3 minutes" · dark, steam, speed6.1%winning
Lunch crowdCombo pricing · fast checkout3.9%arm B

What the data says so far

Late-night angle beats gift angle by +27% CVR at 95% confidence (n=1,204). Suggestion: shift TikTok creative to the 3-minute hook and point its ads at the late-night page.
Promote winnerKeep testing 1 more weekOpen in Studio

Audience × page matrix

Cozy pageSpeed pageCombo page
Late-night5.2%6.1%3.1%
Lunch2.8%4.0%3.9%
Gift4.8%3.3%2.9%

9 · Agent visibility studio

optimize for the LLMs, not just the humans (AEO)

More and more customers arrive via ChatGPT-class agents. This studio checks that your ads, pages and business facts are machine-legible — structured data, agent-crawlable offers, consistent prices — and shows what an agent actually says about you.

ramen.bar

Agent-ready score82/100
Product schema ✓Price + availability ✓llms.txt ✓FAQ for agents ✓Menu schema ✗ missingHours in 3 sources conflictGift page not crawlable ✗

What ChatGPT tells a customer today

"Ramen Bar (ramen.bar) sells at-home ramen kits from $24… I couldn't confirm gift wrapping or delivery to your area."

Missing gift-page crawlability and menu schema = the agent hedges, and hedges lose sales.

Fix queue · applied by workflow

FixImpactStatus
Add Menu schema to /menu+6 ptsREADY
Unify hours across 3 listings+4 ptsREADY
Expose gift page to agents+8 ptsAPPLIED

Adjust anything

These six mocks are the working reference for the v4 implementation card. Tell me what to change — density, colors, a component you hate, a surface missing — and it updates the system before a line of product code is written.

✓ queued — I'll fold it into the system