In Andrew Warner's interview, Impeccable creator Paul Bakaus takes a familiar AI-built site and tries to make it specific to the people who will use it. His point is not that beige pages or stacked cards are forbidden. It is that a model will repeat familiar defaults unless someone gives it a product brief, a design direction, and concrete feedback. Impeccable packages that work into commands, visual exploration, and a browser feedback loop for coding agents.
Impeccable is an open-source design companion for agent-built websites. In the video, Paul uses it to critique a page, explore competing visual worlds, then refine the actual local site with Live picks, annotations, variations, and inserted sections. The useful outcome is a reviewable design process, not a guaranteed conversion lift or a replacement for user testing.
Watch Paul Bakaus Demonstrate Impeccable
Credit: The interview and website demonstration are by Andrew Warner and Paul Bakaus. The episode is presented by Zapier. Paul created Impeccable; its repository and project site are the primary sources for current installation and features.
Three Parts of the Demonstration
1. Critique the page
What Paul shows
He asks Impeccable to critique a site, section, or page. The review calls out weak hierarchy, unclear copy, repetitive AI patterns, and other design issues, then suggests where to focus. The current repository distinguishes critique for UX judgment from audit for technical checks such as accessibility, responsiveness, and performance.
How to use it well
Tell the agent who the visitor is and what they came to do before accepting a score. Treat automated findings as hypotheses to inspect, not proof that a human-made site is good or an AI-made site is bad.
2. Explore visual worlds
What Paul shows
Impeccable asks for context, proposes distinct directions, explains their visual rules, and lets him choose one to build. Paul calls these directions “worlds.” The current project records durable product context in PRODUCT.md and can document an existing visual system in DESIGN.md.
How to use it well
Compare directions for the same audience and task. Keep brand constraints explicit, ask why a direction fits, and choose a concept for its clarity and usefulness, not only for its novelty.
3. Refine in Live mode
What Paul shows
A picker sits over the development site. He selects an element, steers its design, edits copy, draws annotations, requests variants, accepts a result, or inserts a new section. The feedback returns to the coding agent, which writes the accepted change into the local codebase.
How to use it well
Work against a trusted local checkout and review the generated diff. The project explicitly says Live is not a way to edit a deployed production site. Test every accepted variant in the browser and with keyboard input before shipping.
A Better Design Loop for One Real Page
- Write the product truth. Name the primary visitor, their job, the page's purpose, essential content, and brand constraints. A restaurant page for parents booking a meal has a different hierarchy from a kids' reading app aimed at children; Paul makes this distinction at 12:36.
- Inspect the first draft. Ask for critique of one page, then verify the findings yourself. Pick two or three issues that interfere with the main task: unclear action, weak hierarchy, missing contact details, or unreadable mobile copy.
- Explore directions before polishing. Use visual worlds to compare different concepts. Keep a before copy and record why one direction was selected.
- Change the real page in small passes. Use Live to select a specific element, add a comment or annotation, compare variants, and accept only the one that solves the stated problem. Paul's demo includes direct copy edits, voice input, and an inserted case-study section; these are interaction options, not required steps.
- Verify function, not just style. At 17:51, Andrew values a restaurant example because opening hours, phone, booking, and directions are immediately findable. Check the page at mobile and desktop sizes, then test links, forms, text overflow, contrast, keyboard order, and load behavior.
The current repository also describes a deterministic design detector that flags recurring implementation patterns without an LLM call. Paul shows its Rust-based code and agent hooks at 15:45. A fast rule is a useful warning; it cannot decide whether an unusual choice serves the audience.
Install Carefully, and Separate Free Software From Model Cost
The project's current installation instructions recommend running npx impeccable install from the project root, selecting the agent and project or global scope, then starting with /impeccable init where that command is supported. The installer has provider-specific integrations for tools including Claude Code, Codex, and Cursor. Read what it will add to your workspace, particularly hooks, and start on a branch or copy of the project. Codex users should follow the repo's hook-trust instructions after installation rather than assuming the hook is active.
The repository is Apache-2.0 licensed. Paul says the open-source tool will remain free and describes a separate cloud pull-request review beta and possible Pro tier in the interview. Those are creator plans, not features to promise buyers today. Your coding-agent subscription, model usage, and any image generation used for visual exploration can still have costs. The repo also documents a world-catalog request and optional anonymous selection telemetry; review the telemetry section before using it with a client project.
What the Before-and-After Does Not Prove
Paul's examples show a more deliberate visual direction and a more direct way to give feedback. They do not measure conversion, usability, or accessibility improvements. A “design health” score is not a user study. Compare the same real task before and after: can visitors find the right information, understand the offer, complete the form, or contact the business without friction?
Live mode works through a local development server or local static HTML, not a published production site's source. The project's README warns against weakening browser security or production CSP to make Live work remotely. It also notes that local validation hooks can run with your user permissions, so review scripts in an unfamiliar repository before using the feature. Keep client secrets and unpublished material within the agreed boundary.
Jump to a Demonstration
| Time | Segment | Time | Segment |
|---|---|---|---|
| 00:00 | Impeccable introduction | 00:25 | Common AI website patterns |
| 00:54 | Visual worlds | 02:06 | Design-system choices |
| 04:21 | Impeccable Live | 05:06 | Editing beyond annotations |
| 06:18 | Fonts, colors, and styles | 06:45 | Element picker |
| 07:21 | Copy and voice input | 07:48 | Draw annotations |
| 08:18 | Generate variants | 09:27 | Insert a section |
| 10:12 | Agent collaboration | 10:48 | Design critique |
| 12:36 | Design for the audience | 13:21 | Free tool and future plans |
| 15:09 | Repo and agent adapters | 15:45 | Rust detector |
| 16:30 | Install in a project | 16:48 | Before and after |
| 17:24 | Design gallery | 17:51 | Functional design |
Business Idea: A Site Clarity Sprint
A practical offer is not “make your site less AI.” It is a focused review of one page whose visitors cannot quickly find what they need. Use the interview's critique and live-feedback loop to scope a small pilot, then test whether the changes help an actual user complete the task.
One page, one clear task
Design and validate a seven-day site clarity sprint.
Sources and Limits
Video: Andrew Warner interviews Paul Bakaus. Project: GitHub repository, official site and examples, installation, telemetry, and Live-mode limitations. The video is a creator-led demonstration, not an independent benchmark. Commands, adapters, and future pricing can change; use the current repo before installing.