Context Engineering Kit adds judge agents to Claude Code

NeoLabHQ has published Context Engineering Kit, a Claude Code plugin marketplace whose own reliability table rates a one-shot prompt at 1–20% fully accurate results on changesets of more than 20 files. The collection, hosted on Github, ships thirteen plugins covering reflexion loops, spec-driven development, code review, TDD, Kaizen analysis and MCP setup.

At a glance

• Each plugin loads only its own agents, commands and skills into context, and the marketplace command adds the catalogue without pulling any of that material into a session.
• The command /do-and-judge reaches 60% on changesets of 10–20 files at 1.5x to 3x token overhead, while /do-in-steps reaches 71% on the same bracket at 3x to 5x.
• Reliability figures come from more than a year of real development use on production projects, and span approaches from a single prompt up to specification writing with human review before implementation.

Token cost and output reliability trade against each other, and the kit's contribution is to price that trade explicitly rather than leave it to intuition. The numbers are self-reported and not independently benchmarked, so they read as a vendor's operating experience rather than a measurement. What appears to matter more is the shape of the curve: accuracy collapses as changesets grow under one-shot prompting, and the remedies on offer are all forms of context isolation.

One-shot prompting is rated 60–80% on three-file changes and 1–20% past 20 files

The table grades eight approaches against four changeset sizes. A one-shot prompt is rated at 60–80% fully accurate results for one to three changed files, 30–50% for four to ten, 5–30% for 10 to 20, and 1–20% beyond 20 files, at no token overhead.

Adding /reflect lifts the same four brackets to 68–91%, 49–71%, 13–41% and 1–30% for 1k to 3k extra tokens. Pairing it with /memorize, which curates insights into CLAUDE.md, is rated 79–87%, 60–79%, 34–42% and 5–30% for 2k to 5k.

At the top of the table, /plan-task followed by human review and /implement-task is rated 99% on the three smaller brackets and 95% on changes of more than 20 files, at 5x to 35x tokens. Brainstorming ahead of the same flow, without human review, is rated 95%, 95%, 90% and 80% at 5x to 20x.

Spec-Driven Development routes work through eight named agents

The plugin splits into /add-task, /plan-task and /implement-task, with specifications based on the arc42 documentation standard adjusted for model capabilities. Planning runs researcher, code-explorer, business-analyst, software-architect and tech-lead agents, while implementation runs a developer per step, a code-reviewer at the end of each phase, and tech-writer.

Version 3.1.0 embedded DDD and SOLID rules in the developer agent and added the code-reviewer, which applies functional and OOP rules together with Muda waste analysis. NeoLabHQ says the plugin produced working code aligned with the initial prompt in 100% of its tests on production projects.

The plugin also carries MAKER, a reliability pattern from the paper Solving a Million-Step LLM Task with Zero Errors, which relies on clean-state agent launches and filesystem-based memory instead of accumulated context. Quality gates use LLM-as-Judge scoring against predefined rubrics at each planning and implementation step.

The FPF plugin loads a 600k-token specification into a Sonnet[1m] subagent

First Principles Framework implements Anatoly Levenchuk's methodology through an abduction, deduction and induction cycle: generate three to five competing hypotheses, verify them against project constraints, then validate with evidence and compute trust scores. The core specification runs to roughly 600k tokens, so it is loaded into a subagent on Sonnet[1m], which the documentation warns can consume token limits quickly.

Other plugins cover narrower ground. Reflexion cites the Self-Refine and Reflexion papers, which report 8–21% output quality gains across seven tasks, plus Agentic Context Engineering at 10.6% over baselines. Review ships bug-hunter, code-quality-reviewer, contracts-reviewer, historical-context-reviewer, security-auditor and test-coverage-reviewer agents, with four more used for change triage.

Where per-plugin selection is missing

Gemini CLI and Antigravity CLI install the whole bundle of skills and agents, because neither supports per-plugin selection, so unneeded files have to be deleted afterwards. The npx skills route allows picking individual skills but does not carry subagents. The automatic reflection hook, triggered by the word reflect in a prompt, requires bun, though the /reflect command itself does not.