research-sop

name	research-sop
description	Research methodology using Claude Code + Perplexity. Use when conducting competitive research, market analysis, person/lead research (pre-call prep, "who is this person", background checks), or any external research that requires verification. Core principle - Unknown is better than bullshit.
triggers	["research alternatives","compare * vs *","evaluate competitors","find options for","what are the alternatives","market research","competitive analysis","deep research","Perplexity","search who is","research this person","look up this lead","who is this","background check","before the call","sales call prep"]
user-invocable	false

Research SOP Skill

Standard Operating Procedure for conducting research using Claude Code and Perplexity.

Core Principle: "Unknown is Better Than Bullshit"

NEVER fill in data you haven't verified. Mark unverified fields as "Unknown" rather than guessing or assuming.

• ❌ Bad: "Audio Transcription: No" (assumed because research didn't mention it)
• ✅ Good: "Audio Transcription: Unknown" (honest - we didn't verify)

Reasoning: False data causes bad decisions. "Unknown" signals a gap that can be filled later with actual research.

---

Tool Comparison: Claude Code vs Perplexity

Claude Code - Best For:

Strength	Use Case
Codebase analysis	Extracting current technical capabilities from YOUR repos
Context retention	Multi-step tasks where context builds over time
File operations	Reading, writing, organizing research outputs
Data processing	Parsing research results, populating spreadsheets
Tool orchestration	Combining multiple tools (rclone, bash, APIs)
Verification	Cross-checking claims against actual code/docs

Limitations:

• Knowledge cutoff (can't access real-time web data reliably)
• WebFetch/WebSearch are limited compared to dedicated search
• Can hallucinate if asked to "research" without actual sources

Perplexity - Best For:

Strength	Use Case
Deep web research	Finding current info about competitors, markets, trends
Citation tracking	Provides sources for claims (verify them!)
Real-time data	Pricing, funding rounds, recent announcements
Broad exploration	"Who else does X?" type queries
Multi-source synthesis	Combining info from many websites

Limitations:

• No access to YOUR codebase or private data
• Can still hallucinate or misinterpret sources
• Citations need verification - not always accurate
• Lacks context about your specific product/situation

---

Research Workflow SOP

Phase 1: Define What You Know (Claude Code)

Before external research, scan your own repos to establish ground truth:

Scan [your repos] to extract current technical state:
- What's implemented vs planned
- Actual capabilities (not marketing claims)
- Architecture and technical decisions

This prevents drift from reality and gives Perplexity accurate context.

Phase 2: External Research (Perplexity)

Craft prompts that:

1. Provide your context - What you're building, your capabilities
2. Ask specific questions - Not "tell me about X" but "does X have feature Y?"
3. Demand sources - "Cite documentation or say 'Not documented'"
4. List exclusions - What's NOT relevant to avoid noise

Prompt template:

Research [topic] for [your context].

## My Current State
[Paste output from Phase 1]

## Specific Questions
For each [entity], verify:
1. [Specific capability] - cite source
2. [Specific capability] - cite source

## Rules
- If not explicitly documented, say "Not documented"
- Do not assume or infer
- Provide URLs for all claims

Phase 3: Verify & Populate (Claude Code)

Take Perplexity output and:

1. Cross-check claims against cited sources if critical
2. Mark unverified as "Unknown" - don't propagate assumptions
3. Populate structured storage (spreadsheet, database, etc.)
4. Note sources for future verification

---

Red Flags: When Research is Going Wrong

Red Flag	Problem	Fix
All cells filled, no "Unknown"	Likely assumptions	Audit each cell for evidence
Same value for all rows	Copy-paste without verification	Check each entity individually
No sources cited	Claims are unverified	Ask for citations or mark Unknown
Research contradicts your code	Either research is wrong or code changed	Verify against actual implementation
"Seems like" / "Probably"	Speculation not fact	Mark as Unknown
Blog post on someone's site attributed as their product	Content ≠ ownership	Check repo ownership, author field, first-person vs third-person language. Never persist unverified attributions to sheets/registry. Verify first, write second. When challenged, verify IMMEDIATELY.

---

Data Quality Tiers

When populating research data, categorize confidence:

Tier	Meaning	Action
Verified	Confirmed from official docs/code	Use as fact
Reported	From credible source (press, reviews)	Note source, use cautiously
Inferred	Reasonable assumption, no direct evidence	Mark as "Unknown" or note assumption
Unknown	No evidence either way	Leave blank or mark "Unknown"

---

Reusable Research Artifacts

After research, save:

1. Raw Perplexity outputs - For future reference
2. Structured data - Spreadsheet/database with sources
3. Research prompts - So you can re-run with fresh data
4. Assumptions log - What you assumed and why (to verify later)

Location convention:

• Project-specific: [project]/docs/research/
• Competitor research prompts: [project]/docs/competitor-research-prompt.md

---

Person / Lead Research

Researching a specific person (sales lead, contact, candidate) is a distinct workflow from product/market research. Follow this procedure.

Identity Verification Rule

Never assume a search result IS the target person. Common names yield multiple people. A LinkedIn/ZoomInfo/RocketReach profile is Unmatched until corroborated.

Evidence Type	Strength	Example
Unique identifier match	Strong	Same email handle, same phone, same company + role from your source
Photo match	Strong	Profile photo matches known photo (WhatsApp, email avatar, Facebook)
Self-referencing content	Strong	Person's own post/bio confirms details from your source
Name + country + industry	Weak	Common names have many matches — NOT sufficient alone
Name + job title	Weak	Job titles change, multiple people share titles

• ❌ Bad: "LinkedIn shows X as CEO at Y" (assumed match, no corroboration)
• ✅ Good: "Found LinkedIn profile for 'X, CEO at Y' — Unmatched to our lead. No unique identifier overlap."

Tool Selection: Perplexity First

Person research = external entity research → Perplexity is the primary tool.

WebSearch/WebFetch are insufficient:

• Limited result depth, can't synthesize across sources
• LinkedIn blocks WebFetch (HTTP 999), ZoomInfo/RocketReach return 403
• Facebook posts behind login walls

Execution: Use chrome-agent to run queries via Perplexity UI at https://www.perplexity.ai/. Enable Pro Search if available.

Person Research Prompt Template

Research the person "[Full Name]" ([name in other scripts if applicable]) from [Country].

Known facts (from direct communication):
- [Phone/email/handle]
- [Context: where they were found, what they said]

Specific questions (cite sources for each):
1. Full name, location, current role/company
2. Professional background (previous companies, roles)
3. Education
4. Relevant domain experience ([VR/AI/whatever is relevant])
5. Social media profiles (LinkedIn, GitHub, Facebook, Twitter)
6. Any funding, press mentions, or public presence for their projects
7. [Project-specific questions]

Rules:
- If not explicitly documented, say "Not documented"
- Do not assume or infer
- Provide URLs for all claims
- Note: there may be multiple people with this name — distinguish between them

Output Format

Present with confidence tiers from the start. Two tables:

Table 1: Verified (from direct communication) | Data Point | Value | Source |

Table 2: External Research (requires identity match) | Data Point | Value | Confidence | Match Evidence | Source |

Confidence tiers: Verified / Reported / Unmatched / Unknown. "Unmatched" = found data about someone with this name, can't confirm it's the same person.

End every brief with:

1. What you can confirm (direct sources only)
2. What remains Unknown
3. Recommended questions for the call/meeting to fill gaps

---

Quick Reference

Use Claude Code when:

• Analyzing your own code/repos
• Processing and organizing research outputs
• Need multi-step workflows with context
• Populating spreadsheets/databases

Use Perplexity when:

• Researching external entities (competitors, markets)
• Need current/real-time information
• Want cited sources
• Broad exploration of a topic

Always:

• Mark unverified data as "Unknown"
• Note sources for claims
• Re-verify before major decisions
• Update research periodically (quarterly minimum for competitors)