Context Engineering Kit is a marketplace of Context engineering plugins for AI coding agents, published by NeoLabHQ and distributed from a public GitHub repository under the GPL-3.0 licence. It describes itself as a hand-crafted collection of advanced context engineering techniques and patterns with a minimal token footprint, aimed at improving the quality and predictability of an agent’s results, and it targets Claude Code, OpenCode, Cursor, Antigravity and other agents.
The kit’s contents come from two places. The project says the marketplace is based on prompts its own developers have used daily over a long period, supplemented by plugins derived from benchmarked papers and from other projects it judges to be high quality. It publishes a separate list of the research behind that second group, covering refinement loops, memory integration and curation, principle-based critique, evaluation patterns, multi-agent debate, structured exploration, step-by-step reasoning and verification, and hallucination reduction.
What is distributed is prompt material rather than executable software: skills, slash commands, rules and sub-agent definitions that an agent loads into its own context. The project states that its skills follow the agentskills.io specification, and that the specification template used by its Spec-Driven Development plugin is based on the arc42 documentation standard, adjusted for what an LLM can work with.
Capabilities
The kit is installed as a marketplace rather than as a single bundle, and the project treats that
granularity as one of its design goals — each plugin loads only its own agents, commands and
skills, so a user installs just the ones they need. In Claude Code the marketplace is added with
/plugin marketplace add NeoLabHQ/context-engineering-kit, which makes the plugins available
without loading anything into context, and each plugin is then installed individually, for example
/plugin install reflexion@NeoLabHQ/context-engineering-kit.
Other agents are supported with less granularity, which the project states plainly rather than glossing over:
- Gemini CLI installs the repository as an extension, and Antigravity CLI installs from the
repository’s
antigravity/folder. Both take every plugin’s skills and agents as a single bundle; the project notes that neither CLI supports per-plugin selection, and suggests deleting the unwanted skills after installation. - Cursor, Codex, OpenCode and others install through
npx skills add, where individual skills can be picked. Because each of those providers uses its own agent format and that installer does not support sub-agents, the project says this route does not give the full experience. - OpenSkills is offered as a further alternative.
The plugins themselves fall into a few groups. Reflexion supplies /reflect, /memorize and
/critique for feedback and refinement loops, together with a hook that runs /reflect
automatically when the word “reflect” appears in a prompt — the hook requires bun, though the
commands themselves do not. Spec-Driven Development and Subagent-Driven Development implement the
methodologies described in Spec-driven development and
Subagent-driven development. Review provides code and pull-request review through
six specialised agents — bug-hunter, code-quality-reviewer, contracts-reviewer,
historical-context-reviewer, security-auditor and test-coverage-reviewer — with impact and
confidence filtering, and the project positions it as an open-source alternative to CodeRabbit that
can also run in GitHub Actions. The remainder cover Git operations, test-driven development,
Domain-Driven Development rules for Clean Architecture and SOLID, the
First Principles Framework, Kaizen-style root-cause analysis, authoring of an
agent’s own commands and skills, documentation, language-specific rules, and setup of Model Context
Protocol servers.
Adoption & Ecosystem
The project publishes a comparison of how much scaffolding an agent is given against how reliably it produces fully accurate results, running from a bare one-shot prompt through reflection, judge sub-agents and per-file-group execution up to a written specification with human review. Its shape is the argument: reliability rises along that progression while token overhead rises with it, from none for a one-shot prompt to a multiple of the baseline for the specification-driven route, and the gap widens as the number of changed files grows. The figures are the project’s own, described as based on more than a year of real development usage on production projects rather than on an independent benchmark, and the accompanying claims for the Spec-Driven Development plugin — that it produced working code in every case its team tested — are self-reported in the same way.
The project frames its three reliability-oriented plugins as complementary rather than competing, the choice between them being a trade of reliability against token cost, and recommends starting with Subagent-Driven Development and Spec-Driven Development. Two of its other projects are presented as companions: Agent Sandbox, a development sandbox image for agents based on Microsoft’s official devcontainers images, and Agent Eslint Config, an ESLint configuration intended to push agents toward low-complexity, readable code. Both are said to work independently of the kit.