Deep Research — Multi-Hop Recursive Intelligence

Takes a topic through iterative, recursive research across multiple configurable sources. Produces structured intelligence reports in .research/ — not coding briefs (that's context-research).

Flow

/cks:research → configure → plan queries → execute (recursive) → synthesize → save

Mode Detection

Parse $ARGUMENTS to determine mode:

Argument Pattern	Mode	Behavior
"topic"	Topic	Deep dive on a technology, concept, or domain
--competitive "domain"	Competitive Intel	Competitor analysis with comparison matrix
--eval "tech A vs tech B"	Tech Evaluation	Side-by-side technology comparison
--depth shallow|medium|deep	Depth override	Override config default
--refresh	Force refresh	Re-research even if report exists
No arguments	Interactive	Ask what to research

Configuration

Read .research/config.md if it exists. Parse YAML frontmatter for settings.

If no config file exists, create it with defaults on first run:

---
# Source priority: ordered list of research tools to try
# Available: perplexity, context7, firecrawl, websearch, webfetch, huggingface, ahref, mintlify
sources:
  - perplexity
  - context7
  - firecrawl
  - websearch
  - webfetch
  # - huggingface    # Academic papers (uncomment to enable)
  # - ahref          # SEO/competitive data (uncomment to enable)
  # - mintlify       # Documentation search (uncomment to enable)

# Skip sources that aren't available or error (default: true)
skip-unavailable: true

# Depth settings
depth:
  default: medium       # shallow=1 hop, medium=2 hops, deep=3+ hops
  max-queries: 20       # Total queries across all sources per research session
  max-sub-topics: 5     # Max sub-topics to explore per hop

# Output settings
output:
  max-lines: 500        # Max lines per report section
  include-sources: true # Include source URLs and citations
  include-confidence: true # Include confidence scores per finding
---

# Deep Research Configuration

Edit the frontmatter above to customize research behavior per project.

## Source Reference

| Source | Best For | Requires |
|--------|----------|----------|
| perplexity | Broad market/tech research with citations | `PERPLEXITY_API_KEY` in `.env.local` |
| context7 | Library/framework documentation | Context7 MCP connected |
| firecrawl | Scraping specific documentation sites | Firecrawl MCP connected |
| websearch | General web search | WebSearch tool available |
| webfetch | Fetching specific URLs | WebFetch tool available |
| huggingface | Academic papers, ML research | HuggingFace MCP connected |
| ahref | SEO data, competitive domain analysis | aHref MCP connected |
| mintlify | Documentation search | Mintlify MCP connected |

Re-run Check

Before starting, check if .research/{slug}/report.md exists:

• If exists AND no --refresh → show existing report date, ask: "Research found (dated {date}). Use existing, refresh, or new topic?"
  • Use existing: display the report
  • Refresh: archive old report to .research/{slug}/archive/{date}/, re-run
  • New topic: start fresh with different topic
• If not exists → fresh run

Full Flow Execution

When /cks:research is invoked:

1. Parse arguments → determine mode + topic + depth + flags
2. Load config → read .research/config.md (create if missing)
3. Re-run check (above)
4. Validate sources → check which configured sources are available
   • For perplexity: check PERPLEXITY_API_KEY in env
   • For MCP sources: attempt a lightweight call to verify connectivity
   • Report: "Sources available: {list}. Unavailable: {list} (will be skipped)"
5. Route to workflow by mode:
   • Topic → Read workflow: workflows/research-loop.md
   • Competitive Intel → Read workflow: workflows/competitive-intel.md
   • Tech Evaluation → Read workflow: workflows/tech-eval.md

Depth Levels

Level	Hops	Description	Use Case
shallow	1	Single round of queries, no recursion	Quick background check
medium	2	Initial queries + one round of follow-up on discovered sub-topics	Most research tasks
deep	3+	Full recursive investigation, follows citations, cross-references	Strategic decisions, competitive intel

Slugification

Consistent slug generation for directory names:

Topic: "Next.js vs Remix for SaaS" → slug: "nextjs-vs-remix-for-saas"
Topic: "Stripe subscriptions API" → slug: "stripe-subscriptions-api"

Lowercase, spaces→hyphens, strip special chars, collapse multiple hyphens.

Output Artifacts

File	Purpose
.research/{slug}/report.md	Synthesized findings with confidence scores
.research/{slug}/sources.md	All sources with URLs, dates, and reliability notes
.research/{slug}/matrix.md	Comparison matrix (competitive-intel and tech-eval modes only)
.research/{slug}/raw/	Raw query results for later refresh
.research/config.md	Source priority + settings (user-editable)

Integration with Other Skills

Skill	How Deep Research Helps
kickstart/	Replace shallow Perplexity research with Skill(skill="research", args="--depth medium \"{topic}\"")
monetize/	Use competitive-intel mode for competitor analysis
prd/ autonomous	Use tech-eval mode when choosing implementation approaches
context-research/	Complementary — context-research is for quick coding briefs, deep-research is for strategic intelligence

Error Handling

Failure	Behavior
No sources available	Halt, explain which sources failed and how to configure them
Perplexity rate limit	Retry once after 5s, then skip and note gap
MCP source unavailable	Skip if skip-unavailable: true, warn if false
All queries fail for a hop	Save partial results, note "incomplete" in report
Depth budget exhausted	Stop recursion, synthesize what's available

Environment Variables

Variable	Required By	Source
PERPLEXITY_API_KEY	perplexity source	.env.local

Other sources use MCP connections (configured separately).

Reference Files

File	When to Read
references/source-adapters.md	When executing queries — contains the API patterns for each source

Customization

This skill ships with opinionated defaults. Review and adapt to your needs:

• Source priority: Which research tools to try — edit .research/config.md
• Confidence thresholds: When to flag low-confidence findings — edit thresholds in SKILL.md
• Max hops: Depth of recursive investigation (default varies by mode) — edit SKILL.md
• Output format: Report structure and sections — edit SKILL.md
• allowed-tools: Currently Read, Write, Grep, Glob, WebSearch, WebFetch, Agent. Add tools if needed.
• model: Currently sonnet. Remove to use your default model.