Enabling Non‑Technical Users to Build Custom Algorithmic Feeds on AT Protocol via Natural Language

At the recent Atmosphere conference, Bluesky unveiled Attie, an AI‑driven assistant that translates plain‑text commands into custom feed algorithms. This approach targets users who lack programming expertise, allowing them to shape content streams through natural language prompts while remaining within the AT Protocol ecosystem.

Technical Solution

The core engine couples Claude LLM inference with a prompt‑translation layer that maps natural language into feed‑definition schemas. A session manager maintains context, enabling iterative refinement without re‑entering prior parameters. The resulting schema feeds directly into the AT Protocol API, producing a live, queryable feed for the authenticated user.

Implementation relies on containerized microservices orchestrated by Kubernetes, ensuring each component scales independently. Message queues buffer user requests, while caching layers store intermediate feed graphs to reduce latency. All data exchanges adhere to the AT Protocol JSON‑LD format, preserving compatibility across the broader ecosystem.

Architecture Overview

The system is divided into three logical zones: front‑end gateway, processing core, and protocol adapter. The gateway authenticates users via their Atmosphere credentials and forwards sanitized commands to the core, while authentication checks ensure only authorized actions proceed. Each zone communicates using mutual TLS, guaranteeing integrity and confidentiality throughout the pipeline.

The protocol adapter translates generated schemas into AT Protocol feed‑creation calls, handling pagination, sorting, and filter parameters. This layer also logs transaction IDs and status codes for observability, enabling rapid troubleshooting and performance tuning.

Prompt Parsing Engine

Natural language inputs pass through a tokenizer that extracts intent, entities, and constraints. A rule‑based mapper then aligns these elements with predefined feed primitives such as source selection, keyword filters, and ranking weights.

Edge cases trigger a fallback dialog that asks clarifying questions, ensuring the generated schema matches user expectations. The engine logs each transformation step, providing auditability and facilitating future model fine‑tuning.

Feed Generation Pipeline

Once a schema is validated, the pipeline queries the AT Protocol event stream for matching posts, applies ranking algorithms defined by the user, and assembles an ordered list. The list is stored in a temporary cache keyed to the session, enabling rapid retrieval for UI rendering.

Periodic background jobs recompute feeds to incorporate new content, respecting user‑defined refresh intervals. This design balances freshness with compute cost, delivering a responsive experience without overwhelming the network.

Security and Authentication

Attie inherits the AT Protocols decentralized identity model, using DID tokens for proof of user ownership. Each request is signed with the users private key, and the backend performs signature verification before processing.

Sensitive operations, such as feed deletion or permission changes, require multifactor verification via email or authenticator apps. All logs are encrypted at rest, and detailed audit trails are retained for compliance monitoring.