// the find
NeoLabHQ/context-engineering-kit
Hand-crafted Claude Code Skills focused on improving agent results quality. Compatible with OpenCode, Cursor, Antigravity, Gemini CLI, and others. Includes CodeRabbit open-source alternative.
A Markdown-first marketplace of agent plugins for Claude Code, with skills and agents for OpenCode, Cursor and others. It bundles slash commands, sub-agent definitions and rules for planning, reviewing, testing and reflecting on agent output. It is aimed at developers who already use Claude Code daily and want structured workflows like spec-driven development or multi-agent PR review, not a single prompt tweak.
Installs are granular: each Claude Code plugin loads only its own agents and skills, so context cost is paid per plugin. The review plugin runs several specialised agents and filters findings by impact and confidence, and it has a GitHub Action, so it works in CI as well as interactively. The spec-driven plugin uses an arc42-derived spec template with judge-based quality gates between planning and implementation, and the repo's own .specs directory shows the workflow being used on the project itself. Skills follow the agentskills.io format, which in principle keeps the content from being locked to one vendor's tooling.
The reliability table is the headline claim and the weakest part. The percentages are internal estimates from the team's own usage, with no task set, sample size or method, and the 'scientifically proven' framing rests on the Self-Refine and Reflexion papers, which test different setups. Support is uneven: Gemini CLI and Antigravity install every plugin as one bundle, and npx skills drops subagents, so the full experience is effectively Claude Code only. The surface is large, with many overlapping commands such as do-and-judge, do-in-steps, do-competitively and judge-with-debate, and the README gives little guidance beyond two recommended starting points. The review plugin is described as higher quality than CodeRabbit with no comparison behind the claim, and the FPF plugin loads a roughly 600k-token spec into a subagent, which is expensive to run.