oborchers/using-deep-research

Deep Research Methodology

Structured deep research transforms ad-hoc web searches into a repeatable, hallucination-resistant research pipeline. Without deliberate structure, research agents gravitate toward the first sources found, fail to verify claims, and produce confident reports built on unreliable foundations.

This plugin provides 4 methodology skills and the /deep-research:research command for orchestrated multi-agent research sessions.

How to Access Skills

Use the Skill tool to invoke any skill by name. When invoked, follow the skill's guidance directly.

Principle Skills

| Skill | Triggers On | |-------|-------------| | deep-research:research-methodology | Starting any research task — query analysis, decomposition strategies, effort scaling, dynamic replanning, stopping criteria | | deep-research:source-evaluation | Evaluating sources — credibility ranking (T1-T6 tiers), multi-provider search strategy, SEO spam detection, domain-specific source selection | | deep-research:hallucination-prevention | Any research output — hallucination taxonomy, citation verification rules, circuit breaker patterns, confidence scoring, cascading prevention | | deep-research:synthesis-and-reporting | Combining findings — deduplication, conflict resolution, narrative construction, citation formatting, report quality assessment |

When to Invoke Skills

Invoke a skill when there is even a small chance the work touches one of these areas:

• Starting research: Load research-methodology to plan decomposition and effort scaling
• Searching the web: Load source-evaluation to assess what you find
• Writing any claim: Load hallucination-prevention to verify before stating
• Combining findings: Load synthesis-and-reporting to merge and cite properly

The /deep-research:research Command

For full orchestrated research sessions, use /deep-research:research. The command:

1. Checks web access permissions (one-time setup per project)
2. Analyzes the research query — if too vague, asks 2-3 clarifying questions
3. Decomposes into subtopics based on query complexity (not a fixed number)
4. Spawns parallel research-worker agents (Sonnet) — each writes findings with a Verifiable Claims Table
5. Spawns parallel research-verifier agents (Sonnet) — each re-fetches sources and checks claims independently
6. Dispatches a research-synthesizer agent (Opus) — applies corrections, merges findings, writes final document with Confidence Assessment
7. Preserves intermediate docs and verification reports for traceability

The Three Meta-Principles

All research skills rest on three foundations:

1. Every claim needs a source — No unsourced assertions. If it cannot be cited, it cannot be stated as fact. Flag uncertainty explicitly.

2. Source quality determines output quality — 57% of research errors originate in early retrieval. Front-load high-quality sources. Prefer primary sources (T1-T2) over secondary sources.

3. Verify before synthesizing — Treat each agent's output as untrusted input. Cross-reference claims between sources. Use deterministic validation where possible.