Release
Your knowledge base stays focused — even when it's big
PRD generation now keeps knowledge-base context tight and prioritised instead of pushing every uploaded document into every draft — so generation stays fast and on-point, and you control exactly which documents a draft uses.
If your team had uploaded a stack of reference docs to the Brain, every PRD draft was quietly carrying all of them — full text, every time. With a few large files that meant a slower, more expensive, less focused generation, where the genuinely relevant document competed for the model's attention with everything else you'd ever uploaded.
What changed
Generation now works to a knowledge-base budget. It includes your most recent documents in full and stops once it has enough context, instead of dumping the entire library into the prompt. When there's more than fits, you'll see a short note in the output saying the knowledge base was trimmed and how many documents made it in — no silent truncation pretending everything was read.
The result: drafts stay fast and stay on-topic, even as your Brain's knowledge base grows.
You're still in control
When a specific draft should lean on specific documents, select them on the Context step before you generate. A hand-picked selection always wins over the automatic budget — so the one spec you actually need is never the one that gets trimmed. See Set up your Brain for where uploads and per-PRD selection live.
Under the hood
We also added per-call token and cache telemetry to the admin dashboard, so we can see exactly where generation spends tokens and how much of each prompt is being served from cache. It's the groundwork for keeping Thinkr's AI fast and inexpensive as the product grows — measured, not guessed.
What stays true
Nothing about quality changed here — the same grounded, structured PRDs, the same honesty rules. This is about spending context where it counts.