More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Improve UI: Evidence-Based Interface Fixes

    1

    Audit an interface against its own design system instead of taste: prove each finding three ways, try to falsify it, and hand off self-contained fix plans. Strictly read-only on product source.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Improve UI: Evidence-Based Interface Fixes

    Improve UI: Evidence-Based Interface Fixes

    Example session with this skill installed

    Our checkout flow "feels off" — support keeps hearing that buttons and prices look inconsistent between the cart, payment, and confirmation steps. Can you audit it and tell us what to fix? We don't want opinions; we want to know what's actually wrong.

    • Read your context and instructions
    • Compiled the improve ui

    AUDIT — checkout flow (cart → payment → confirmation): one task family, one layout shell.

    Design language: the token file and component library govern this flow. Documented decisions that matter here: prices always render through the Money component; one primary action per step; step headers use the Title variant.

    Seven candidates collected. Through the proof gate:

    Finding 1 (High) — The confirmation step formats its total by hand. Contract: the Money component spec names checkout totals explicitly. Runtime: the confirmation template builds the string inline and never imports Money, producing a different currency/decimal layout than cart and payment. Correction: render the total through Money; no other change.

    Finding 2 (Medium) — The payment step carries two primary-styled actions ("Pay now" and "Save card"). Contract: the documented one-primary-per-view rule, contradicted within the same task. Correction: demote "Save card" to the secondary variant, following the exemplar in the account settings flow.

    Killed at the gate: "the price font feels bigger on payment" — traced values resolve to the same token; the impression is surrounding whitespace, which no contract governs. "Cart button should match the marketing site" — a different governing system, out of scope. Three more candidates died in falsification: a deliberate A/B variant, a documented exception, and a duplicate of Finding 1.

    Improve first: Finding 1. Every buyer sees the order total in two different formats inside a single checkout, and the fix is a one-component swap. A plan for Finding 1 is ready on request; Finding 2 is reported without a plan until you choose it — that sequencing is deliberate.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most UI feedback is taste wearing a lab coat. This skill runs an interface audit where findings are earned, not asserted. It starts by bounding one surface and reconstructing the design system that actually governs it — documentation verified against what renders, tokens followed through aliases to their definitions. Candidates are collected cheaply, then each must pass a three-proof gate: a binding contract (a documented rule, or a direct contradiction in user-facing presentation within the same task), runtime proof that the cited value actually reaches the surface, and a single correction the evidence requires — ambiguity kills the candidate. Survivors are attacked in a falsification pass, assigned confidence from a rubric, and reported at three findings maximum, ordered by confidence, user impact, reach, and fix cost. "No supported findings" is a respected outcome, stated plainly. The auditor never touches product source: the output is the report, plus — only on request — one self-contained implementation plan per finding, written so another engineer or agent can execute it without the audit conversation.

    • Presentation only: functional bugs, performance, data wiring, and behavior changes are out of scope unless a documented design contract governs them; accessibility findings are excluded unless explicitly requested.
    • Findings are capped at three per audit by design — a surface with more real problems gets sequential audits, not a longer list.
    • Where no design documentation exists, contracts can only come from contradictions within the same task, so findings will be fewer and narrower. That's the method working, not failing.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 24 days ago

    Permissions

    No special permissions declared or detected

    Listed24 days ago
    Updated21 days ago

    What's inside

    Frequently Asked Questions

    Every finding must cite a binding rule, prove the offending value actually reaches the rendered surface, and name the single correction the evidence supports — then survive an explicit attempt to falsify it. Opinions, frequency counts, and "best practice" claims don't pass the gate.

    The audit reconstructs what governs the surface from tokens and shared components, and contracts can still come from direct contradictions inside the same task. Expect fewer, narrower findings — and the report says so, instead of inventing rules to have something to report.

    Separation of duties. The auditor's value is evidence an executor can trust; mixing in edits would compromise both. Each plan is written to be executed blind — exact files, current and required values, and a "done when" check — so handoff loses nothing.

    New in Frontend & Web Apps