Structured Rubric

A structured rubric is the encoded judgment your brand runs at AI speed. It consists of five to seven concrete criteria each scored one to five with a mandatory one line reason a pass threshold like average four or higher and nothing below three and strict JSON output. Feed it an AI candidate rendered as image text or code plus the prompt and Claude or GPT returns structured data you can sort filter and audit. This is the LLM as judge layer in the eval pyramid. It sits after lint and visual diff and before the human taste review. It kills the mediocre stuff cheaply so your senior designers only see the top fifty candidates instead of ten thousand. The rubric gets versioned like code because it is code. You test it monthly against real conversion data from shipped surfaces. Criteria that predict clicks or saves get weighted higher. Criteria that do not get revised or dropped. In 2023 ML teams built these to ship models faster than humans could review. Designers stole the pattern in 2025 and by 2026 the teams without one look like they are still designing with their thumbs in Slack threads.

A structured rubric is not a brand deck. It is not a paragraph that says capture our innovative spirit with delightful micro animations. Those documents sit unread in drives while AI ships off brand garbage by the terabyte. It is not a loose set of principles you interpret differently each time you write a prompt. It is not theater where you ask an LLM to score without a threshold so every candidate gets a four and you learn nothing. It is not written once and forgotten. A rubric that never updates becomes a fossil of last years opinions. It is not the full picture of taste. It cannot make the final call when three options all score four point three and you need to pick which one feels alive. That stays human at the top of the pyramid. It is not a replacement for senior eyes. It is the lever that multiplies them.

Here is the exact voice rubric we ship with ClaudeBrainy. The prompt starts with Score the copy 1 to 5 per criterion. Deliver one line reason per score. Then the criteria hit hard. One Lead first. Does the first sentence directly answer the users question or is it still warming up. Two Concrete. Does it name real products actual numbers specific moves or does it hide behind abstractions. Three Voice match. Does the tone rhythm and vocabulary match the brand profile we documented in our brand system last October. Four No filler. Does every sentence carry weight or is it there to sound smart. Five No banned constructions. Zero em dashes zero AI slop adjectives like seamless or groundbreaking zero hedging phrases like sort of or in some cases. The pass line sits at average four point zero or higher with no individual criterion below three. The output must be parseable JSON containing scores as an object reasons as an object and a top level pass boolean. We ran this on sixteen hundred AI drafted support articles in March 2026. Total cost twelve dollars. Runtime under four minutes. It rejected seventy four percent. The survivors went to human review where we shipped twenty two pieces that performed twenty eight percent better on read time than the old vibe based process. One failure scored one point eight on concrete for saying cutting edge technology instead of specifying how our integration with Cursor reduced bug rates from one in fifty to one in four hundred. Another tanked voice match for using passive constructions Stripe banned in their 2024 style guide. We built a parallel rubric for layout craft. Criteria include adherence to the eight column grid never sixteen pixel padding violations clear visual hierarchy that works at thumbnail size whitespace rhythm that follows our four pixel baseline and zero orphan elements that float without relationship. This one processed four thousand variants generated by v0 Lovable and Claude Artifacts in a single batch. It flagged an entire cohort that ignored our coral amber cream cyan palette in favor of generic blues. Linear runs a version of this on every writing brief. Their mid level designers no longer debate tone in threads. The rubric decides and the data from user engagement closes the loop every thirty days. Vercel does the same for Geist updates. Every new component variant gets scored before it touches production. Stripe applies identical logic to design system contributions. The output feels like magic brand consistency. The reality is a living document that gets tuned when metrics show whitespace criteria predict engagement better than hierarchy ones. We even built a color rubric last quarter that scores palette adherence contrast without sterility and semantic token usage. It caught three separate v0 runs that invented new hex values instead of pulling from our token set.

Deploy a structured rubric the day your team starts drowning in AI output. That threshold hit most teams in late 2025. Place it squarely in the middle of your eval stack so lint catches the broken grids visual diff catches the unintended token changes and the rubric catches the off brand voice and craft failures. Hook it to your existing Claude API calls or use Anthropics eval framework if you want versioning out of the box. Run it on every generated candidate that survives the cheap layers. Sort the results. Route anything that passes to a short human queue. Then feed conversion metrics back in. If high scoring concrete copy drives more saves raise that criteria weight next month. Use it when you want to free senior designers from drudgery and point their taste at real decisions. The teams at Stripe and Linear who adopted early now operate eval systems that cover surface area three times larger than their 2024 headcount with tighter quality. Do not bother with a structured rubric if you generate under two hundred candidates per week. The overhead is not worth it and your existing LGTM loop still works. Do not use one if your brand lead and design lead did not sit in the room to define the criteria together. Solo authored rubrics encode one persons blind spots and produce brittle enforcement. Never launch without a hard pass threshold. Without it the scores become meaningless numbers that fail to filter. Never skip the one line reason field. You need those to spot patterns in failures and improve the rubric itself. Never treat it as set it and forget it. A rubric that does not evolve with your brand and your data becomes a liability that locks in yesterday is taste. The entire system fails if you automate the human layer above it. Taste calls stay human on purpose.

A structured rubric is how you stop vibing your brand and start engineering it at the speed of AI.