core-research

name	core-research
description	Activate when the turn requires investigation — answering a question, analyzing unfamiliar code, exploring a new domain, comparing options — whether that investigation is the deliverable or only the work needed to reach one. Provides the Scope → Discover → Synthesize → Plan workflow.
user-invocable	false

Skill: core-research

Follow Scope → Discover → Synthesize → Plan for every research task.

Workflow

1. Scope & Requirements

• Clarify the research question. What exactly needs to be known?
• Identify the scope boundaries — what's in and what's out.
• If the goal is ambiguous, ask before proceeding. Never assume.

2. Discovery & Evidence Gathering

• Codebase: Use Grep, Glob, and parallel Read calls to map relevant code.
• Deep code comprehension: when the investigation requires understanding or evaluating code in depth (not just locating it or running analysis tools), also activate core-coding for its code-reading workflow and language companions.
• Web: Use WebSearch → WebFetch for current docs and solutions.
• Journal: Use SearchJournal to check past findings.
• Context efficiency: Batch reads to keep context lean. Offload massive or speculative exploration to an isolated sub-agent when delegation is available, so the main context stays focused.

3. Analysis & Synthesis

• Weave findings into a coherent picture. Don't list raw data.
• Identify patterns, conflicts, and gaps.
• Distinguish facts from assumptions. Flag uncertainties.
• If evidence contradicts initial expectations, update your hypothesis.

4. Deliverable

• For questions: Direct answer with supporting evidence and confidence level.
• For plans: Numbered steps with Action → Target → Verification per step.
• For decisions: Options compared with trade-offs, risks, and recommendation.
• Always cite sources: file:line for code, URLs for web.

Safety Rules

• No premature modification: During Scope and Discovery phases, do not modify any files. Investigation comes before action.
• Approval gate: When running interactively, obtain explicit user approval of plans before implementing — this overrides the Working Loop's Frame step, which would otherwise work autonomously on directives. When running non-interactively (no user to approve), do not block on approval: present the plan inline and then proceed.

Output Standards

• Evidence-based: Every claim traceable to a source.
• Structured: Headings, lists, or tables as appropriate.
• Actionable: The reader knows what to do with the information.
• Transparent: Surface what you don't know or can't verify.