research

Systematic Research

When to Use This Skill

Use when:

• User explicitly requests research or investigation (/ct:research command)
• Questions that require exploring multiple sources
• Synthesising information from diverse sources
• Current events or real-time information beyond your knowledge cutoff
• Multiple competing claims need verification and synthesis

When NOT to Use This Skill

• Simple factual questions within your knowledge base
• User explicitly wants just your existing knowledge
• Code implementation tasks (use TDD, debugging skills)
• Questions answerable by reading local codebase files
• Pure brainstorming or ideation (use brainstorming skill)

Boundaries (DO NOT)

• Do NOT write implementation code or code examples — this skill gathers external knowledge; hand off to implementation skills/agents.
• Do NOT make architectural or technology decisions for the user — present findings neutrally; defer the choice.
• Do NOT skip workflow steps or citation requirements — all four phases are mandatory regardless of query simplicity.
• Do NOT provide "best guess" recommendations when authoritative sources are unavailable — move the gap to Open Questions instead.
• Do NOT use general-knowledge statements as findings — every factual claim must cite an authoritative source or be moved to Open Questions.

Hard Caps

• Maximum sources per topic: 25 — stop collecting beyond this unless the user explicitly extends.
• Maximum depth levels: 2 — do not recursively follow citation trails beyond 2 levels.
• Maximum report length: 8000 words — state this at the top of the report. If research exceeds this, split into sections and deliver in parts.

The Four-Phase Research Process

HARD GATE - Research Phase Completion: → For each phase (Discovery → Investigation → Synthesis → Reporting): Check: Have I completed this phase's deliverable? No → Complete it before moving to next phase. Skipping is not permitted.

Phase 1: Discovery

• Restate the question in one sentence; list unknowns; flag blocking assumptions.
• Choose depth: quick (1 authoritative source), standard (2-3 sources cross-referenced), deep (multi-hop with contradictions resolved), exhaustive (survey with per-claim confidence).
• Formulate queries.
• Execute broad searches in parallel (via tavily-cli skill and/or WebSearch).
• Assess sources: credibility, bias, recency.

Trigger check: If the first batch of searches returns 0 results from authoritative sources (primary/official docs, peer-reviewed, reputable news), reformulate queries with different terms before proceeding to Phase 2.

Phase 2: Investigation

• Apply multi-hop reasoning: entity expansion, temporal progression, conceptual deepening, causal chains.
• Route extractions:
  • Static HTML → WebFetch, or via the tavily-cli skill (invoked through the Skill tool).
  • JavaScript-heavy sites → via the agent-browser skill (invoked through the Skill tool).
  • Library/framework docs → via the context7-cli skill (invoked through the Skill tool).
  • Local files/codebase → Read, Grep.
• Collect evidence with citations (URL + title + access date).

Trigger check: If any source contradicts another source, state the contradiction in the report and prefer the more authoritative (primary source > secondary > tertiary). Do not silently pick one.

Phase 3: Synthesis

• Connect information across sources.
• Identify patterns, contradictions, gaps.
• Distinguish facts from interpretation.

HARD GATE - Citation-or-gap rule: → About to include a factual claim in the synthesis → Is this from an authoritative source I can cite? Yes → Include with citation. No → Am I tempted to say "probably" / "likely" / "typically"? → STOP. Move to Open Questions section instead. Explicitly state the gap rather than filling with general knowledge.

Trigger check: If you have <2 sources for a claim flagged as "key finding", downgrade it to a single-source observation in the report, or move it to Open Questions.

Phase 4: Reporting

• Present findings with confidence levels.
• Cite every factual claim.
• Acknowledge limitations.

Report template:

Report cap: 8000 words. If research exceeds this, split into sections and deliver in parts.

## Goal
<one-line restatement of the question>

## Findings
- <bullet with inline citation [1]>
- <bullet with inline citation [2]>

## Sources
| # | URL | Title | Credibility | Note |
|---|-----|-------|-------------|------|

## Open questions / suggested follow-up
- <explicit gap or uncertainty>

Parallel Agent Fan-Out

→ Research task with multiple search threads → Can these run independently (no dependency between results)? Yes → Dispatch Agent subagents in ONE message, one per independent facet, maximum 10 subagents per batch. If >10 facets, batch sequentially in groups of 10 (wait for each batch before dispatching the next). No (true dependency) → Run sequentially.

Each subagent investigates one facet of the question, then you synthesize their findings.

Tool Orchestration

• WebSearch / WebFetch — primary for web search and static-page extraction.
• tavily-cli skill (invoked through the Skill tool) — richer web search, extract, crawl, deep research.
• context7-cli skill (invoked through the Skill tool) — library/framework documentation.
• agent-browser skill (invoked through the Skill tool) — JavaScript-heavy sites.
• Agent — parallel research threads for independent questions.
• Read, Grep — local files/codebase.

Quality Standards

Information: cross-reference claims across >=2 sources; prefer primary sources (official docs, post-mortems, peer-reviewed) over secondary (news coverage) over tertiary (blog summaries); note publication dates; flag bias explicitly.

Synthesis: distinguish facts from interpretation; mark speculation with the word "speculation"; acknowledge contradictions by name; assign confidence levels (high / medium / low) per claim; show reasoning chains.

Anti-Patterns

Anti-Pattern	Correct Approach
Cherry-pick	Don't stop at first source — seek contradicting views
Sequential-crawl	Don't search one-by-one — batch queries in parallel (fan-out block above)
Speculation-pass	Don't say "probably" — cite sources or move to Open Questions
Citation-skip	Don't claim without sources — every fact needs attribution
Confidence-fake	Don't present uncertainty as fact — state confidence levels
Replan-resist	Don't push through dead ends — reformulate queries when 0 authoritative results