qengine

A Claude Code plugin for software engineering built on the harness engineering model — every component is a Markdown file with YAML frontmatter that the Claude Code harness loads at session start. No build step, no runtime, just structured prompting with strict division of labor.

Install

Add the marketplace and install the plugin from the GitHub repo:

/plugin marketplace add phuquocchamp/qengine
/plugin install qengine@qengine

Or via the CLI:

claude plugin marketplace add phuquocchamp/qengine
claude plugin install qengine@qengine

What's inside

Skills

Skills are invoked with a slash command inside Claude Code.

Skill	Command	Purpose
deep-research	/deep-research <topic>	Multi-round source-heavy investigation. Spawns parallel researcher subagents, drafts a full report, adds inline citations, and produces a verified brief with provenance tracking.
explain-code	/explain-code <file | symbol>	Explains code with an analogy, a Mermaid diagram, a step-by-step walkthrough, and a gotcha. Saves the explanation as a dated Markdown file.

Agents

Subagents used by the research pipeline. They do not invoke each other — only the deep-research skill orchestrates them.

Agent	Role
researcher	Gathers primary evidence across web, papers, repos, and docs. Writes an evidence table with verifiable URLs.
writer	Turns research notes into a structured draft. Preserves caveats; adds no citations.
verifier	Anchors every factual claim to a source, verifies each URL, and builds the final Sources section.
reviewer	Acts as a skeptical peer reviewer or adversarial auditor. Produces a structured review with inline annotations.

The research pipeline

deep-research is the orchestrator. It fans out to parallel researchers, writes the draft itself (deliberately bypassing writer to keep evidence traceability tight), then hands off to verifier and reviewer.

---
config:
  theme: default
---
flowchart LR
  A[deep-research SKILL] --> B[researcher x N parallel]
  B --> C[Lead writes draft]
  C --> D[verifier adds citations]
  D --> E[reviewer audits]
  E --> F[final brief + provenance]

Loading

All artifacts for a run live under a dated folder:

./docs/deep-research/<yyyy-mm-dd>-<slug>/
├── .plans/<slug>.md
├── .drafts/<slug>-draft.md
└── no-NN-<descriptive-name>.md

The no-NN- zero-padded prefix encodes execution order across researcher rounds and post-processing steps.

Design principles

• Harness engineering model. All behavior is defined in Markdown with YAML frontmatter. The Claude Code harness is the runtime — no external dependencies.
• URL or it didn't happen. No source is cited without a direct, checkable URL.
• Preserve uncertainty. Inferences are labeled as inferences; contradictions are surfaced, not smoothed.
• Strict division of labor. Researchers gather, the lead drafts, the verifier cites, the reviewer audits. No agent does two jobs.
• File-based handoff. Subagents write to files and return lightweight references so the parent's context stays small.

Layout

.claude-plugin/      # plugin manifest and marketplace catalog
agents/              # subagent definitions
skills/              # skill definitions
  deep-research/     # multi-agent research orchestrator
  explain-code/      # code explanation with diagrams
CLAUDE.md            # editing conventions and invariants

Requirements

• Claude Code CLI or desktop app
• Claude Max or API subscription (the deep-research skill runs on Opus by default)