Field Report: Coding in the Age of AI with Cursor (drezil.de)

This field report walks through a pragmatic workflow for coding with Cursor — a VSCode fork that tightly integrates LLMs with the editor by injecting files into prompts, exposing filesystem and editor state, executing shell commands, and letting tool-aware models read linters, language servers and tests. The author emphasizes that Cursor isn’t magic: it can accelerate planning and implementation but is constrained by token-window limits, query costs, data‑security/legal exposure and a practical tendency to generate redundant code when it can’t reliably discover existing utilities. The report reframes success as strong project management and rigorous specifications rather than letting an agent “figure it out” on its own. Technically, the author built rule-driven guardrails: rules are Markdown files with YAML frontmatter (description, globs, alwaysApply) that inject behavior and context into LLM prompts. Example rules include an “always apply” project-conventions rule (e.g., run tests with poetry run pytest, no bash idioms), a PRD generator that asks up to five clarifying questions and only emits /tasks/<feature>/PRD.md after a “go” signal, and detailed specification-writing guidelines (sections, error handling, testing requirements). The practical workflow: discover need → investigate and produce PRD → pin a spec → check spec-compliance → generate tasks → then implement. Implication: to safely scale agentic IDEs you need explicit specs, tooling constraints, and human-in-the-loop decision gates to avoid technical debt and privacy/legal pitfalls.