Field notes on Context-Driven Development, the knowledge-graph engine, and what it actually takes to make AI write production code.
An AI agent with no context writes plausible code that quietly breaks something three files away. Draft gives your assistant a spec, a plan, a live code graph, and a three-stage review — turning a fast guesser into a disciplined engineer. Free, local, open source — a five-minute tour of the whole workflow.
Read postVector search returns the chunks that look like your query. Draft's knowledge wiki returns the concepts that actually matter — by reasoning over a navigable table of contents instead of running a similarity search over embeddings. No vector DB, no chunking, no cloud — and a retrieval path you can audit.
Read postOn two real codebases — a 17K-LOC Rust service and a 221K-LOC polyglot platform — grep is often faster than Draft's graph engine. It's also wrong far more often (10% vs 100% hotspot precision), and it costs an AI agent up to 108× more tokens to reach the correct answer. In an agent loop, latency is the axis that matters least.
Read postLoading every MCP tool schema upfront costs 50–140K tokens and breaks selection accuracy past 30–50 visible tools. Four agents, same architectural pattern, four incompatible implementations — and one outlier still paying the full bill on every turn. A 2026 field comparison.
Read postA real impact-mode run on a hotspot before a single line gets written. 47 downstream files, 12 tests, 3 config templates — and a refactor plan reshaped before /draft:implement starts. Three loops, one shared graph: pre-change risk, post-completion memory, cross-track overlap.
/draft:decompose WalkthroughA real run on a 14-file track: eight modules, a circular dependency caught and broken before any code, and an LLD with contracts for the two high-complexity modules. The decomposition skill isn't a planner — it's an architectural guardrail.
Read postMost "AI for code" tools improve context by sending more of your code to a third party. Draft's graph engine does the opposite: a local, live-queried knowledge graph, powered by codebase-memory-mcp and installed alongside the plugin. No MCP server, no SaaS, no cloud, no API key, no leaked code.
Read postAmazon's Kiro and Draft both solve the same problem: AI coding without structure is chaos. The similarities end there. Bug hunting, ACID enforcement, three-stage review, architecture discovery, IDE portability, and a $0 price tag — the honest comparison.
Read postWant every post in your feed reader? Subscribe via RSS. Want the long-form deep dive? Read the Draft Book.