Polish user-facing product surfaces such as pricing pages, plan cards, feature pages, grant/application packets, navigation rows, empty states, model/settings pickers, provider import/export prompts, provider logos, App Store/TestFlight metadata, icons, categories, onboarding, and marketing UI by reading real reference products and the product's enforced source of truth before writing copy, visuals, budgets, or interaction structure. For research grant packets, keep funder-required research intent distinct from product/prototype framing.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Polish user-facing product surfaces such as pricing pages, plan cards, feature pages, grant/application packets, navigation rows, empty states, model/settings pickers, provider import/export prompts, provider logos, App Store/TestFlight metadata, icons, categories, onboarding, and marketing UI by reading real reference products and the product's enforced source of truth before writing copy, visuals, budgets, or interaction structure. For research grant packets, keep funder-required research intent distinct from product/prototype framing.
Product Surface Polish
Use When
Editing pricing pages, plan cards, upgrade/paywall screens, feature lists, in-app feature pages, navigation rows, empty states, model pickers, auth/sign-in provider buttons, settings copy, App Store metadata, TestFlight metadata, app icons, product categories, onboarding, or marketing UI.
A product/settings/auth surface names third-party providers or tools and needs their logos, brand marks, icons, or provider-choice affordances.
A product/settings surface uses another AI's import/export prompt, dialog, or onboarding flow as a reference for the product's own prompt or flow.
The user points at reference products and asks for the surface to feel similar, bigger, cleaner, more premium, simpler, or less awkward.
Editing a grant, application, or submission packet where the funder's required materials, page limits, applicant type, budget rules, line-item cost assumptions, or optional award resources define the artifact.
Editing a research grant/application packet where the research question, evaluation method, and experimental apparatus must stay distinct from product, pilot, or prototype deliverables.
The reference arrives through a help article, FAQ, video, screenshot, or live browser page, and the product surface depends on its layout, copy, or interaction pattern.
The user names a real app, device, screen, or artifact contract for a product surface, such as an existing app Upgrade screen or a specific simulator.
Product or proposal copy/visuals need to preserve a user-corrected architecture, runtime split, or reference shape.
Copy includes internal terms, vague feature claims, filler benefits, or numbers that may drift from enforced plan limits.
Near miss: do not load this for backend-only billing logic unless a user-visible product surface is being changed.
Procedure
Name the exact surface and audience: store listing, home-screen app identity, pricing page, locked paywall, onboarding, model picker, grant packet, or in-app settings. If the user named a real app path, device, screen, or artifact contract, include that in the surface contract before acting.
For grant/application/submission packets, read the official application page and write the artifact contract before drafting: required attachments, page limits, budget requirements, applicant category, deadline, submission email/form, and selection criteria. Before finalizing a budget, build a source-backed cost model for the material lines: parts or services needed, quantities, unit prices or defensible ranges, vendor/source class, and why the compressed one-page line item matches the project shape. Keep sidecar award resources such as credits, access, support, or provider APIs out of the proposal thesis and section structure unless the application asks for them or the user explicitly asks to feature them. For research grants, name the research claim, evaluation method, and apparatus separately before editing: the device, prototype, or pilot is the instrument unless the official requirements make it the deliverable. Lead deliverables with research artifacts such as a study protocol, evaluation suite, dataset, results, apparatus, or technical report; do not make a working product or pilot the headline when the funder asks for research.
Read the real references the user named, plus the closest direct competitors when the user asks "how do others do it." For browser, help-article, FAQ, video, or screenshot references, save and read durable artifacts before editing: full-page screenshots or frames, article images and metadata, FAQ text, and the relevant DOM or visible text. Extract the pattern in concrete terms: naming formula, navigation placement, icon choice, subtitle placement, empty versus populated state, primary creation affordance, example prompts or rows, card/list hierarchy, feature taxonomy, spacing, and responsive behavior.
Classify each reference's role before implementing: verbatim source to copy, inspiration to synthesize from, or interaction/component behavior to match. For behavior to match, write the interaction contract before changing code: what the empty state invites, where input lives, which examples appear, how the layout changes once data exists, and what confirms success or failure. Use one success confirmation; when the pressed control visibly changes to Copied, a checkmark, or another done state, do not add a duplicate success toast. Copy provider prompt text verbatim only when the user asks for copying; when provider prompts are examples of how another app extracts/imports data, write one product-owned prompt shaped by the pattern instead. When the user corrects the product architecture, runtime location, or reference object, propagate that correction through headings, captions, image prompts, diagrams, and verification steps before generating more artifacts.
Read the product source of truth before writing copy: plan constants, enforced quotas, billing config, app metadata, bundle settings, feature gates, and current screenshots. A pricing claim must trace to code or config; store metadata must trace to App Store Connect or bundle settings.
Gotchas
A reference list is not proof you read references. Summarize the pattern you observed before changing the product.
Prior browser access is not durable reference evidence. If the UI decision depends on screenshots, FAQ text, article images, or a live page, capture those artifacts before the first edit.
Copying a reference's labels, icon, or route is not enough. For feature pages, the creation path, empty-state examples, placement of the input, and transition to the populated state are part of the product pattern.
Public names, subtitles, category labels, and pricing bullets are user psychology surfaces. They should not expose internal architecture, debugging shorthand, or implementation limits unless that is the sellable benefit.
Model/tier picker labels and descriptions are product positioning copy. Read peer pickers first, then map internal model capability into short task-language; do not ship backend/provider metadata or vague filler like "needs more care."
If the product has plan constants, copy should be generated from or checked against them so the page cannot promise more or less than the runtime enforces.
Store review metadata is downstream of the in-app purchase path. A screenshot that matches pricing is not proof the app can select or buy that plan.
Preview flags, generated concept renders, and proposal shorthand are proxies. Replace them with the named real surface or corrected architecture before claiming the product artifact is fixed.
Award-side resources are not automatic proposal sections. A funder mentioning credits, model access, support, or an API in the award announcement does not mean the application asks for a credits plan, API plan, or provider-centered thesis.
A one-page grant budget is not permission to invent a round hardware or services number. Do the BOM or cost-model work first, then compress it into the attachment.
A research apparatus is not automatically the funded product. Do not answer that the funder wants research and then list a working app, device, or pilot as the first deliverable; put the study/evaluation artifact first and describe the device as apparatus.
Provider-choice rows and sign-in buttons are brand-recognition surfaces. A GPT label, single letter, or hand-drawn approximation reads as unfinished next to real Claude, Gemini, ChatGPT, Google, Apple, or Grok marks.
Prompt references can be examples, not requested copy. When the user says another app's prompt is inspiration, preserve the product's own voice and data contract instead of either inventing an unrelated prompt or pasting the provider's exact wording.
A provider import dialog is still an app dialog. Start from the repo's default dialog component; a custom wide modal or cloned competitor layout is scope creep unless the user asked for that surface.
Success toasts are not free polish. If the control itself flips to , checked, done, or saved, that is the confirmation; keep toast/banner feedback for errors or background outcomes that otherwise have no visible state.
Verification
Positive trigger: "Read ChatGPT, Claude, Manus, and Perplexity pricing and fix our pricing cards."
Positive trigger: "Read Plaud/Granola/Pocket and fix our TestFlight name, category, and icon."
Positive trigger: "Fix the model picker wording; how do Claude or ChatGPT describe these tiers?"
Positive trigger: "The settings import row needs ChatGPT, Claude, Gemini, and Grok logos instead of placeholder text."
Positive trigger: "The Continue with Google sign-in button needs the real Google logo, not a letter."
Positive trigger: "Use Claude and Gemini's import prompts as inspiration for our memory import dialog."
Positive trigger: "Fix the memory import dialog copy affordance; the Copy button already turns into Copied."
Positive trigger: "Before uploading App Store subscription screenshots, check that the iOS paywall actually opens the selected plan."
Positive trigger: "Put the iOS paywall into Companion Staging iPhone 17 Pro and open Upgrade in the existing app."
Positive trigger: "Read the ChatGPT Scheduled help article, all images, and FAQ before making our empty state match."
Positive trigger: "Update this grant packet so the headset is only capture/output and the AI runtime is off-head/API."
Positive trigger: "Fix this Thinking Machines grant packet; Tinker credits are optional and the application wants the interactivity proposal materials."
Positive trigger: "What parts are needed for this grant budget? You did not research the hardware line."
Positive trigger: "This grant is research, so why does the packet sound like I am building a product?"
Near miss: "Change the backend limit for automations from 10 to 20" with no UI/store/pricing surface.
Sources
Internal failure patterns: model/settings copy was changed from intuition instead of reading reference product language; provider-logo UI shipped placeholder text; App Store review screenshots were prepared before proving the paywall path; scheduled-task UX copied naming while missing the creation-first empty state; proposal/grant packets drifted from corrected architecture and required materials; a redundant success toast duplicated an existing copied state.
Skill format and routing rules: skills/skill-creator/references/source-map.md.
For subscription or paywall work, prove the in-app purchase path before store-review artifacts: inspect the paywall code/config, product IDs, selected-plan params, StoreKit/Superwall wiring, and a real simulator/device screenshot or flow. If the user named an existing app, simulator, or screen, install/launch that exact target and navigate the normal app path; a preview flag, standalone fixture, or different booted device is a proxy unless the user explicitly asked for it. Do not upload App Store review screenshots or polish subscription metadata while the app only has static cards or an unproven trigger.
Translate internal capability into user language. Remove engineering phrases and filler such as "tool-heavy," "model context window," "standard usage," "switch anytime," or "add more credits anytime" unless the reference products and product truth make them genuinely meaningful.
For tiered pricing, make progression visible from low to high: more quota, broader capability, clearer support/access. Keep duplicated roll-up lines short, avoid unlimited claims unless the runtime actually enforces no limit, and align cards across desktop and medium-width layouts.
For provider import/export dialogs, use the app's existing standard dialog and controls first. Custom sizing, typography, or cloned competitor chrome needs a concrete reason from the reference pattern and the local design system.
For third-party provider logos, including auth buttons, search local repo assets first, then current web, vendor, and public icon sources before inventing a mark. Prefer official brand kits or established public SVG icon packages; keep the resulting asset local in the app with its source recorded. Text initials or generated marks are temporary implementation scaffolds, not the final product surface, unless no real mark exists after search.
For app-owned icons and store metadata, choose the model/tool intentionally after checking available local/API options. Generate or edit the asset only after the reference pattern is clear, then verify the built app bundle uses the asset and the public metadata changed.
Verify on the actual surface, not a proxy: screenshot or inspect the live page, run typecheck/lint/build, verify responsive wrapping, and for referenced interaction models verify the relevant states that define the pattern, not only that the route renders. For App Store/TestFlight work confirm the processed uploaded build or current App Store Connect fields.
Copied
A generated visual can be worse than a clean existing glyph if the prompt is not grounded in the app identity and peer icon language.
Labels such as "AI headset" can silently reintroduce the wrong architecture after a correction to off-head, phone, pocket, edge, cloud, or API runtime. Scrub titles and captions, not only body paragraphs.
