deep-research

name	deep-research
description	Skill for deep research grounding using Copilot CLI /research for multi-source literature review, prior-art search, and SOTA anchoring. Use before forming hypotheses to anchor claims in real, citation-backed evidence.
user-invocable	true
disable-model-invocation	false

Deep Research Skill

Grounding investigations in real evidence before forming hypotheses.

When to Use

Use this skill when tasked with:

• Literature review or prior-art search
• SOTA anchoring (what's the current best approach?)
• Grounding hypotheses in existing evidence
• Verifying claims before committing them as findings

The Problem This Solves

LLM agents can "recall" information from training data without citations, leading to fabricated or outdated claims in knowledge briefs. The /research command provides citation-backed evidence from live sources, eliminating training-data-only recall as a knowledge source.

How to Use

/research <specific, focused question>

The command searches GitHub repositories + live web sources and returns citation-backed results with source URLs.

Protocol

1. Formulate specific queries — not "machine learning" but "continual learning methods that prevent catastrophic forgetting in transformer models"
2. Run /research for each major topic area
3. Extract citations — record source URLs in strategic-context.md
4. Cross-reference with Semantic Scholar for academic depth (if available)
5. Record all sources in the knowledge brief with inline citations

Problem Positioning Queries

When doing positioning research for the Scout, craft queries that find the FRONTIER, not the textbook:

BAD: "transformer attention mechanisms" GOOD: "2025 2026 advances in structured context encoding for LLMs"

BAD: "multi-agent systems" GOOD: "multi-agent coordination for scientific discovery latest results"

BAD: "optimization" GOOD: "high-dimensional coupled decision optimization with heterogeneous constraints recent papers"

Query design rules:

• Include year markers (2025, 2026) to bias toward recent work
• Include "state of the art" or "latest results" to avoid survey papers
• For the closest-work query, include the specific technique or contribution
• For methodology comparison, include "methodology" or "approach" plus the specific technique
• Do NOT search for textbook-level concepts — search for what CHANGED recently

Three mandatory positioning queries (Scout Phase 0):

1. Frontier: "[field] latest results 2025 2026 state of the art"
2. Sub-problem: "[specific technique/approach] recent advances methodology comparison"
3. Closest work: "[one-sentence investigation description] published results"

Rules

• NEVER rely on training data recall for factual claims — always verify via /research
• If /research returns no relevant results, state "no prior art found" explicitly
• Each claim in the knowledge brief must trace to a /research citation or a Semantic Scholar paper
• Run as many /research queries as needed to ground claims properly

When NOT to Use

• Simple codebase search (use grep/search tools instead)
• Questions answerable from the current repository
• Tasks with tight time budgets where premium request cost matters
• Build/implementation tasks (these don't need literature grounding)