FinishUI / curated workbench

FinishUI Builder Library

A curated, source-reviewed library for building cleaner AI-made apps. Start with one useful resource, understand its limits, and keep the path back to a consistent interface.

Open the free UI-Repair Prompt Kit Browse the library

Free toolkit / no account

UI-Repair Prompt Kit

Eight focused prompts for turning “the UI feels off” into observable checks and agent-ready fix rules. Use them with ChatGPT, Claude, Gemini, or a coding agent. The model still needs your real route, constraints, and human judgment.

8 promptsCopy and adaptNo code upload required

01 / Diagnose

Name the visible drift before fixing it

Review this public UI screenshot or route description. List only observable inconsistencies in typography, color, spacing, repeated controls, interaction states, and mobile behavior. Separate evidence from taste. Do not redesign yet. Return the five highest-impact holds in review order.

02 / Rules

Turn repeated decisions into a small UI system

Convert the approved UI corrections below into a compact rule set for type roles, color tokens, spacing steps, radii, controls, states, and responsive behavior. Reuse existing components where possible. Flag any rule that lacks evidence instead of inventing one.

03 / Fix plan

Give the coding agent a bounded repair order

Create a dependency-ordered UI cleanup plan for this working route. Preserve product behavior and backend logic. Fix shared tokens and components before local overrides. For every task, name the visible problem, the smallest change, the screens affected, and the check that proves it is done.

04 / Mobile

Check the narrow-screen boundary

Inspect this interface at 430 pixels and 320 CSS pixels. Identify overflow, unreadable type, clipped controls, broken hierarchy, unreachable actions, and content that depends on hover. Recommend the smallest responsive correction for each observed issue.

05 / Recheck

Rerun the same checks after the fix

Compare the corrected route with the approved UI rules. Rerun the same named checks under the same viewport and state conditions. Return pass, hold, or not-tested for each check. Do not call the route finished when evidence is missing.

06 / Reference extraction

Extract decisions from a UI reference

Study the approved reference and extract only reusable interface decisions: type roles, color roles, spacing rhythm, component geometry, density, emphasis, state treatment, and responsive behavior. Separate what is visible from what you infer. Do not copy branding, content, or proprietary assets.

07 / Component-state map

Map every state before the agent invents one

Create a component-state map for the controls on this route. Include default, hover, focus-visible, active, selected, disabled, loading, success, warning, error, and empty where relevant. Mark states that are observed, specified, inferred, or not applicable. Do not invent behavior that changes product logic.

08 / Evidence-bound handoff

Hand off the fix without losing the proof

Prepare an evidence-bound implementation handoff. For every approved UI change, include the observed problem, governing rule, affected component or route, exact acceptance check, viewport or state condition, evidence location, and owner approval status. Keep unverified assumptions in a separate hold list.

These prompts help structure diagnosis. They do not prove accessibility conformance, security, conversion lift, or client approval.

Free visual field maps

How a human-approved UI enhancement system actually works.

Save the diagrams, share them with your team, or use them to stop a coding agent from treating a confident output as a finished interface.

Six-stage FinishUI UI enhancement loop from evidence to human approval — Map 01UI Enhancement Loop
Observe, diagnose, decide, implement, recheck, and approve.
Open the full SVG

FinishUI human-in-the-loop architecture connecting owner intent, agents, evidence, and browser verification — Map 02Human-in-the-Loop UI Architecture
Automation carries the work. The owner retains judgment and release authority.
Open the full SVG

FinishUI original diagrams. They describe one practical operating pattern, not a required standard or proof of product quality.

V1 / deliberately small

Tools, Prompts, Skills, Agents, and Repos

Each lane starts with one useful resource. More entries earn a place only after readers click, use, request, or convert.

01Tools

Official source · Direct unpaid link · Last reviewed July 15, 2026

Replit Agent

Best for turning a scoped product idea into a working browser-based prototype without setting up a local development environment first.

Why it earned inclusion: It combines prompting, code access, preview, and deployment in one place, which lowers setup friction for nontraditional technical founders.
Important limitation: Fast generation does not guarantee a coherent design system, correct product logic, accessibility, or production readiness. Credit use also rewards tighter scopes.

Open Replit Agent at its official site

02Prompts

FinishUI original · Free resource · Last reviewed July 15, 2026

UI-Repair Prompt Kit

Best for turning vague interface discomfort into named evidence, shared UI rules, a bounded fix order, and a same-condition recheck.

Why it earned inclusion: It mirrors the diagnosis, system, and verification workflow used throughout FinishUI's public proof.
Important limitation: A prompt cannot see missing product context, negotiate taste with a client, or prove the implementation matches the recommendation.

Use all eight prompts

03Skills

Official repository · Direct unpaid link · Last reviewed July 15, 2026

Vercel Web Design Guidelines Skill

Best for giving a compatible coding agent a reusable design-review workflow instead of repeating the same design instructions in every prompt.

Why it earned inclusion: It comes from Vercel's official agent-skills collection and follows the portable Agent Skills pattern.
Important limitation: Installation compatibility and output quality still depend on the agent, project context, source evidence, and review process.

Vercel Web Design GuidelinesUI review rules for compatible agents. Anthropic Frontend DesignA production frontend direction skill for Claude Code. Vercel React Best PracticesPerformance rules for React and Next.js work. Open Agent Skills installerInstall compatible skills from source repositories.

Compatibility note: review the source before installation. Availability, commands, and agent support can change, and a skill cannot replace project context or human approval.

04Agents

Official documentation · Direct unpaid link · Last reviewed July 15, 2026

OpenAI Agents SDK

Best for building an application that needs agents, tools, guardrails, handoffs, sessions, and tracing without inventing the orchestration layer from scratch.

Why it earned inclusion: The SDK exposes a small set of agent primitives and includes human-in-the-loop and tracing mechanisms.
Important limitation: An agent framework does not decide your product boundaries, security model, approval rights, or whether an agent should take an action.

Read the official Agents SDK documentation

05Repos

Official repository · Direct unpaid link · Last reviewed July 15, 2026

OpenAI Cookbook

Best for learning implementation patterns from runnable examples before asking an agent to invent an integration from memory.

Why it earned inclusion: It is OpenAI's official examples repository and provides concrete notebooks and code for common API patterns.
Important limitation: Examples are starting points. Verify model names, APIs, security, costs, and production behavior against current documentation.

Browse the official OpenAI Cookbook

How recommendations work

Affiliate relationships never decide what earns a place.

Every link is labeled as direct unpaid, owned by FinishUI, referral, affiliate, or partner. At launch, every external V1 recommendation uses a direct unpaid link. If FinishUI later receives compensation, the label will change beside the link before the tracked link appears.

Video learning paths

FinishUI Watch Desk

Founder Watchlist

Six direct learning destinations grouped by the decision they help you make. This is a practical study path, not a popularity ranking or proof that every video is current.

Official course · Direct unpaid link · Last reviewed July 16, 2026

01 / UI foundations

Figma Design for beginners

Build a responsive portfolio site while learning frames, auto layout, components, and prototyping. Best starting point for understanding the design decisions AI builders often flatten.

Open the official Figma course

Expert education channel · Direct unpaid link · Last reviewed July 16, 2026

02 / Visual judgment

Nielsen Norman Group

Use its research-led UX explainers to sharpen the reasoning behind hierarchy, interaction, navigation, and usability decisions instead of copying surface style.

Open the NN/g YouTube channel

Official product channel · Direct unpaid link · Last reviewed July 16, 2026

03 / Design systems

Figma

Follow official tutorials and product lessons on components, variables, auto layout, and shared libraries. Confirm current feature behavior in Figma's documentation.

Open the official Figma YouTube channel

Also useful for responsive site craft: Framer Academy.

Official company channel · Direct unpaid link · Last reviewed July 16, 2026

04 / Agentic coding

Anthropic

Use official Claude and Claude Code demonstrations to understand capabilities and workflows, then verify commands and limits against current product documentation.

Open the official Anthropic YouTube channel

Official product channel · Direct unpaid link · Last reviewed July 16, 2026

05 / AI product shipping

Replit

Study real Agent workflows and product launches when you need the shortest path from a bounded idea to a working browser-based prototype.

Open the official Replit YouTube channel

Prefer a structured course? Start with Replit Learn.

Independent creator channel · Direct unpaid link · Last reviewed July 16, 2026

06 / Responsive craft

Flux Academy

Use selected web design and portfolio lessons to study composition, client presentation, and responsive craft. Treat creator opinions as guidance, not standards.

Open the Flux Academy YouTube channel

FinishUI review standard

A sponsor can buy attention, never the verdict.

Same-condition test

Run the tool on a declared task, route, viewport, and evidence set so readers can understand the result.

Visible limits

Publish what failed, what was not tested, the plan used, and the last-reviewed date.

Material connection beside the claim

Label paid, gifted, affiliate, referral, partner, and direct unpaid relationships where readers see the recommendation.

No paid rankings

Paid promotions remain visibly separate and cannot change editorial scores, comparison order, or negative findings.

Field guides

The workflow behind cleaner AI-made products.

Long-form guides are published one at a time after research, sourcing, and review. The buyer-focused cluster comes first.

PublishedHow to Fix a Vibe-Coded Website Before Client ReviewDiagnose visible drift without redesigning the whole app.

Next guideBest AI-Assisted UI WorkflowA practical sequence from reference and rules to implementation and visual recheck.

ResearchingHow to Build an AI UI Quality PortfolioA no-hype pathway for showing evidence, process, and judgment when learning UI/UX or AI product design.

Applied help

Want FinishUI to diagnose and apply the fixes?

Your app can already work while the UI still feels inconsistent. Send one authorized public route. FinishUI will identify the visible holds and tell you whether the free route is enough or an applied cleanup fits.

Request my free UI scan

The Private Office / one company at a time

Sometimes the UI problem is really an operating problem.

A screen can expose a larger system: disconnected AI tools, unowned workflows, and growth decisions no one can audit. For one qualifying company at a time, the Private Office works at that layer.

No public calendar · No generic discovery calls

01Earned accessPrior FinishUI work or a trusted executive introduction.
02Sponsored mandateA named decision-maker and a cross-functional AI or growth problem.
03Serious intentA real decision window and an approved five-figure budget.

The full qualification brief stays private. Only applications that clear it receive a conversation.

Request the private brief