LGTM Loop

The LGTM Loop is the casual review habit that defined design teams for fifteen years. A designer finishes the mock. Drops the Figma link in Slack. Colleagues reply with LGTM or a thumbs up emoji. The design lead opens the file for thirty seconds before standup. That counts as approval and the asset ships. The system worked because the bottleneck sat at creation not review. One designer produced one hero image or one landing page per week. Human attention matched human output exactly. No rubrics existed. No automation ran. Just quick vibe checks from people who sat in the same room.

The LGTM Loop is not a quality system. It carries no structured criteria. It ignores deterministic failures a machine can catch in milliseconds. It never connects to conversion data that improves the next round. It scales on human hours instead of compute dollars. At AI volume it becomes a coin flip with extra steps. Tools like v0, Cursor, Claude Artifacts and Lovable generate six hundred variants per brief. Thirty briefs equal eighteen thousand candidates. Teams still using Slack threads ship contrast failures, invented hex colors, broken four pixel grids, off brand voice and touch targets under forty four pixels at industrial scale. ML engineers solved the identical problem in 2023 when models outran human review. Designers who refuse to steal that playbook waste their best talent on mechanical errors.

Look at the concrete example of a fintech team modeled after Stripe in 2024 versus 2026. In 2024 the single product designer produced one new checkout flow every sprint. The file went into the design reviews Slack channel. Three engineers typed LGTM. The lead added one comment about button hierarchy. It shipped. Minor contrast issues got fixed in QA two weeks later. No one lost sleep. Early 2025 the team connected Claude and v0 to their brand system. One prompt now generated four hundred checkout variants overnight. Thirty active briefs created twelve thousand candidates daily. The Slack channel turned into an unmanageable dumpster fire. Senior designers burned entire mornings rejecting obvious padding errors and type scale escapes. One principal calculated she personally reviewed over four thousand candidates in February alone. Burnout hit hard. The team finally copied the ML eval playbook. Layer one lint with axe core and a custom Figma token validator killed thirty eight percent of output before any human saw it. Layer two isolated components in Storybook, captured screenshots with Playwright, ran pixel level diffs with Pixelmatch and surfaced changes in Chromatic. Layer three deployed a custom Claude rubric with seven measurable criteria: lead first copy, concrete product names and numbers, exact voice match, zero filler, banned constructions including em dashes and hedging, proper grid adherence and accessible touch targets. The model returned JSON with one to five scores per criterion, one line reasons and a pass fail verdict at four point zero average with no score below three. Only the top fifty candidates reached human eyes. Those reviews became high leverage taste calls between strong options. Conversion data from live A B tests started flowing back monthly. Criteria that actually drove save rates got weighted higher. The rubric became versioned living code audited quarterly. Output consistency improved dramatically. The old Slack LGTM thread now sits archived except for memes about how they once worked. Linear ran the exact same transition on their product writing in Q1 2025. Vercel applied it to Geist typography updates. Both teams report senior designers now spend time on strategy instead of acting as expensive lint brushes.

Use the LGTM Loop only for low volume low stakes work. Solo side projects. Early wireframes for internal tools. Teams of five or fewer who sit together and ship under twenty items per week. It still functions there because human attention matches output. Stop using it the moment AI generation scales. Never apply it to customer facing surfaces at companies moving at real velocity. Stripe hit the wall in late 2024 when AI generated design system variants reached production with subtle regressions. They replaced the LGTM habit inside one sprint. Linear ditched it for copy after their Claude powered variants overwhelmed editors. Both teams built rubrics that encode taste at scale and tied them to production metrics. The rule is brutal. When any designer reviews more than one hundred candidates per week the LGTM Loop has become technical debt. Build the pyramid immediately. Run deterministic lint at the base. Add visual regression on every change. Score survivors with LLM as judge against a concrete rubric. Reserve human taste for the final fifty options and for intentional rule breaks the model cannot make. Close the loop with conversion signals that retrain the rubric every month exactly as Vercel does with Geist and Stripe does with their component library. Teams that make this switch ship stronger work with fewer hours. Their seniors finally operate at leverage instead of scale.

The LGTM Loop collapses at AI scale because human taste does not parallelize but a layered eval stack of lint, diff, judge and taste absolutely does.