Cursor Agentic Toolkit
A structured operating layer for AI-assisted delivery in Cursor: intent → brief → spec → tasks → QA, curated context, external operational memory, and installer profiles—without turning the toolkit repo into your product workspace.
https://github.com/bertom/cursor-agentic-toolkit
What this is
The Cursor Agentic Toolkit is a lightweight way to run Cursor agents like a small, disciplined team: clear intent, traceable decisions, real QA, and explicit human-in-the-loop gates. It started from the same itch as methodology repos like the BHIL toolkit—specification quality is the bottleneck—but focuses on Cursor-native behavior: where rules load, how git stays clean, and how the next session picks up from files instead of chat.
Insight into the repo
The kit is toolkit source: templates, .cursor/rules, guides, and install-runtime.sh you apply into an independent project repo. Daily work stays in the app repo; optional external operational memory (project-ops symlink) holds specs, task packs, and client-facing material without bloating product history. Runtime summaries live under something like .agentic/; full ops detail can live outside when you want hygiene.
Documentation covers layer model, external setup, workflow chain, human tasks, context system, runtime profiles, upgrades with --dry-run, and why Cursor only auto-loads rules from <project-root>/.cursor/rules/—not from nested toolkit folders.
Benefits
- Same pauses a good team takes — Encoded as markdown artifacts so the next session continues, not reinvents.
- Curated, file-based context — Context index and stable ground truth instead of thread-only memory.
- Clean product git — Runtime vs operational memory split;
.gitignoredefaults you can preview. - Honest automation boundaries — Human task lane for things agents must not fake (logins, keys, legal).
- Repeatable install — Profiles, scripted install, dry-run; onboarding is not folklore.