Seven years of Loomery projects have produced a lot of hard-won knowledge on how to run a project. Last year we wrote it up in ‘The Loomery Way’ , a Notion site kept up-to-date with our own standards on how to run discoveries, shape technical strategy, work with AI safely, build a product roadmap and more.
But a static site only helps when people know what information exists and where to look for it.
So we created Bobbin Slack is where our team already asks how things are done and what resources exist, so that’s where Bobbin lives: an agent that has access to all of The Loomery Way and answers questions using our playbook, taking in the Slack thread context. It joins Warp , our product brain, and Heddle , our AI skills library, which together sit at the core of YarnAI , our AI-native method for shipping better products.
How it works You can message Bobbin directly or tag it in a Slack channel. Say an engineer kicking off a new project asks in the channel, ‘How do I design an agent that has access to a knowledge store of whitelisted books and articles?’ Bobbin:
Answers in the open: Bobbin replies in the Slack thread, so everyone in the channel sees the answer, not just the person who asked.Reads the thread: It takes in the conversation so far and its answer fits what's being discussed instead of starting from scratch.Shows its working: Every answer links the Notion pages it drew on. We can check its sources and access any further resources it recommends.Only gives the Loomery answer: Answers are grounded in our tried and tested approach, with past examples.Suggests follow-up steps: Uses the thread context and content from The Loomery Way to suggest logical next steps.In answer to that engineer’s question, Bobbin provides an outline for designing that agent including core guidance on checking if an agent is producing reliable output and productionising agents, as well as links to verified sources included in the cited Loomery Way articles.
For existing Loomers this gives direct access to resources and a reminder on our core principles. It also helps us get new team members up to speed quicker, giving them access to the company's memory from day one, so they can get started running projects.
Building Bobbin In one line: Bobbin is Claude Haiku using retrieval-augmented generation (RAG). It searches The Loomery Way using the Notion MCP and answers only from what it finds.
We chose Haiku because it's the cheapest in the Claude family but still capable enough to answer vague questions. If a question doesn’t directly match an article title, it tries similar keywords until it finds relevant content using the context in Slack to judge this. Haiku’s large context window allows it to read long Slack threads, picking up details from several days or weeks ago and giving better responses.
The current iteration of Bobbin calls AWS Bedrock’s Converse API on the backend, a simple API that calls any model in Bedrock’s catalogue using a universal schema. This gives us the option to swap out the model by simply changing the model ID in the code. In the API, the system prompt sets the rules, like never responding with anything irrelevant and always searching our Notion site, not the model’s training data.
Notion and Slack were added as tools through the API’s toolConfig so the model is aware of both Slack and The Loomery Way and how to search them. Each tool is split out into actions like notion_search and notion_read_page so the model decides what it needs at any point in the loop before responding in Slack. Bobbin collects every article it uses with the link, citing it wherever the article is mentioned.
Currently, Bobbin does one job well: searching The Loomery Way. From the way it's been built this could be extended to connect up to other specialist YarnAI agents too, for example, Jira ticket writing and coding agents. This is only the beginning of how we can use Bobbin in our daily work.
Best practice that answers back Bobbin makes the knowledge that was already there easier to find, and allows us to interact with AI in the tools we already use. Together with Warp and Heddle, it’s how YarnAI, our method for running AI agents across a whole project, makes sure AI gains build up across the team instead of staying with individuals.
If your organisation’s best practice is stuck in heads and wikis, we’d love to help you build your own version of Bobbin, get in touch .
