Hammy is an agentic project management assistant built on the Anthropic API. It interfaces with an Obsidian vault to understand project specifications, milestones, and tasks — and can write new notes, generate PRDs, and manage task lists through a conversational interface.
The Problem
Personal project management across a large, evolving knowledge base is messy. Notes get stale, tasks get scattered, and synthesizing status across multiple projects requires manually reading through dozens of files. Hammy solves this by treating the vault as a live context window.
What I Built
A Python-based RAG pipeline that ingests markdown files from an Obsidian vault, scores and retrieves relevant notes based on natural language queries, and calls the Claude API with that context. A Flask web interface serves a styled chat UI locally. Write operations — creating notes, appending tasks, generating documents — require explicit confirmation before touching the vault.
What I Learned
Retrieval quality is more important than model size for grounded, personal-context applications. Structured frontmatter dramatically improves the signal-to-noise ratio in retrieved chunks. The confirm-before-write pattern is essential when an agent has write access to anything you care about.