Second Dinner Turns Card Prototypes Around in Hours, and Says AI Tests Beat Manual Ones in 81% of Cases
Second Dinner, the studio founded by the creators of Hearthstone, connected a RAG system to a Slack bot. The model first pulls past cards, code and Jira tickets from LanceDB, and only then answers. A designer now gets a mockup for a new card in hours rather than months. QA calls an API instead of spending a week on manual work, and the studio estimates that AI-generated tests beat human-written ones in 81% of cases. The takeaway: an internal RAG pays off where ideas stall in approval cycles.
AI-processed from LanceDB Blog; edited by Hamidun News
Imagine a company with years of code, design documents and tickets, while its neural network knows nothing about them. The game studio Second Dinner, founded in 2018 by Hearthstone creators Gamilton Chu and Ben Broussard, solved this with RAG. This is the name for a setup in which the model searches the company's own data for relevant pieces before answering and relies on them.
The studio stores its data in LanceDB, a vector database that searches documents by meaning rather than exact words. As a result, a mockup for a new game card now takes hours instead of months, and the studio itself estimates that AI-generated tests are better than those written by people in 81% of cases. The studio prepared the case together with the LanceDB team.
What problem the studio had
There were two pain points. First, systems designers constantly come up with new mechanics, but turning a sketch into working code requires gathering engineers, UX designers and testers. That means months of meetings, spent budget, and ideas that never reach testing.
Second, every new game card has to be checked against thousands of interactions with other cards. QA engineers wrote these test suites by hand, and it took a full week. The team was small, so missed edge cases were a real risk.
The studio set out its requirements as follows: a prototype generated from a text prompt, test generation via API, the ability to work with heterogeneous data (from Jira tickets to design documents), and a serverless architecture without infrastructure hassle.
How the solution works
The AI team converted the company's codebase, design documents and Jira tickets into vectors and stored them in LanceDB Cloud. According to the description, setup came down to uploading embeddings (numerical "fingerprints" of text) and creating indexes. The platform handled the rest: scaling, dataset compression, incremental indexing, and full index rebuilds without stopping running services. It took two weeks from getting access to production. The studio explains its choice with a "no-ops" philosophy (minimal manual maintenance) and cost: by its account, LanceDB was 3–5 times cheaper than alternatives. The case does not say which alternatives were compared.
An example for a designer: they write in Slack, for instance, "What would a card that steals mana from the opponent's pool to boost my spells look like?" Through RAG, the system finds similar mechanics from past cards and then automatically assembles a web page mockup. The designer shows it to colleagues and gets feedback without assembling an engineering team or risking the release of an untested feature. According to the studio, as a result designers are running 5 times as many ideas in parallel.
QA engineers get a test set through an API call or a Slack command, and this takes minutes rather than a week. According to the studio's data, automatically generated tests are better than those written by people in 81% of cases, and 20.5% of them are "noticeably better" in formatting or requirement coverage.
What to keep in mind
This case was co-prepared by the customer and the vendor, so the figures should be read with caution. The test comparison methodology is not disclosed: it is unclear who judged "better," by what criteria, on what sample, and how edge cases were counted. The claim about a fivefold increase in parallel ideas is also given without details. Neither the models, the embeddings, nor the search quality evaluation scheme are named. Phrases like "from executors to visionaries" are marketing, not an engineering result.
At the same time, the architecture itself is clear and reproducible. The value does not come from a separate model but from access to the company's own data: past cards, rules and tickets serve as context on which the LLM generates plausible prototypes and tests in the style of a specific project.
Practical takeaway
For teams building internal AI assistants, the case suggests three lessons. First, start with what has already accumulated: code, documents and tickets, rather than an abstract chatbot. Second, the interface should live where people already work: Slack and APIs are more convenient than a separate portal. Third, a managed serverless database spares a small AI team from maintaining indexes, which matters when there are no dedicated DevOps specialists on staff.
If you are building a similar system, agree in advance on how you will measure success. The share of tests that beat manual ones, the time from idea to mockup, and the cost of storage and search are the metrics the studio itself names. But without a transparent methodology, such numbers do not transfer well to other projects, so run your own measurements before scaling up.
- Jun 11, 2026Opponents of AI Regulation Gather in Washington as the 2026 Elections Change the Rules
- Nov 12, 2026"Tochki nad II" Author Replaces Notion and CRM with Agents: 80+ Skills Instead of Services
- Nov 9, 2026Pinky the agent runs a CRM of 1,000+ people: a Second Brain built on md files and git instead of Notion
- May 2, 2026Microsoft launches Copilot Health to analyze data from smartwatches and medical records
Key Questions Answered
What problem the studio had?
There were two pain points. First, systems designers constantly come up with new mechanics, but turning a sketch into working code requires gathering engineers, UX designers and testers. That means months of meetings, spent budget, and ideas that never reach testing. Second, every new game card has to be checked against thousands of interactions with other cards. QA engineers wrote these test suites by hand, and it took a full week. The team was small, so missed edge cases were a real risk.
How the solution works?
The AI team converted the company's codebase, design documents and Jira tickets into vectors and stored them in LanceDB Cloud. According to the description, setup came down to uploading embeddings (numerical "fingerprints" of text) and creating indexes. The platform handled the rest: scaling, dataset compression, incremental indexing, and full index rebuilds without stopping running services. It took two weeks from getting access to production.
What to keep in mind?
This case was co-prepared by the customer and the vendor, so the figures should be read with caution. The test comparison methodology is not disclosed: it is unclear who judged "better," by what criteria, on what sample, and how edge cases were counted. The claim about a fivefold increase in parallel ideas is also given without details. Neither the models, the embeddings, nor the search quality evaluation scheme are named. Phrases like "from executors to visionaries" are marketing, not an engineering result.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.