k-deep-research（深度研究）

K Deep Research v2.0

Universal research methodology for any domain, any topic, any complexity level. Optimized for OpenClaw autonomous agents AND Claude.ai project workflows.

⚠️ CRITICAL: Load Before Researching

When research is requested, you MUST:

1. Read this SKILL.md (you're doing it now — good)
2. Load references/sourcing-strategies.md — WHERE and HOW to search
3. Load domain-relevant references as needed (see Reference Map below)
4. Execute the 7-step workflow
5. Output as Obsidian-ready .md file (YAML frontmatter mandatory)

DO NOT skip this skill and jump to web search. Methodology > raw queries.

Core Research Workflow

Execute in sequence for every investigation:

1. CONTEXT CHECK    → Existing knowledge base / prior research
2. QUERY ELABORATION → Expand scope, plan search strategy
3. MULTI-SOURCE      → Gather from diverse sources (40-80+ for deep)
4. PATTERN ANALYSIS  → Cross-domain recognition, temporal/actor/info flow
5. CREDIBILITY SCORE → 0-10 scale on ALL sources, merit-based
6. SYNTHESIS         → Compile findings preserving contradictions
7. OUTPUT            → Obsidian .md with YAML frontmatter

Research Principles

Institutional Skepticism: Official narratives = data points, not truth claims. Merit-Based Sources: All sources start equal. Evaluate on internal consistency, specificity, predictive accuracy, corroboration potential, incentive analysis, technical coherence. Peer review is not a truth guarantee; institutional rejection is not falsification. Pattern Recognition: Temporal clustering, actor coordination, information flow, anomaly correlation, historical precedent, narrative consistency. Epistemic Humility: Absence of evidence ≠ evidence of absence. BUT systematic patterns of absence ARE informative. Physics First: Technical feasibility analysis before accepting exotic claims. Adversarial Analysis: Cui bono? Suppression signatures? Inversion test (what if the "debunking" is the disinformation)?

Tool Selection Strategy

SearXNG (PRIMARY for sensitive/adversarial research):

• Zero telemetry, aggregates across engines
• Use for: institutional analysis, suppression tracking, contested topics
• Fallback: built-in web_search when SearXNG unavailable

Web Search (general research):

• Current events, academic papers, community discussions
• Non-sensitive technical topics

Context7 MCP (technical documentation):

• Code libraries, frameworks, APIs, SDKs
• Coverage: 30k+ snippets across dev ecosystem
• NOT for: consciousness, legal, historical, institutional topics

Filesystem (existing knowledge):

• Obsidian vault (4000+ files)
• Prior investigation notes, timelines, frameworks

Decision Tree:

Sensitive/adversarial topic?  → SearXNG first
Code/framework/API docs?      → Context7 first
Existing research available?  → Filesystem first
General research?             → Web search
Always:                       → Multi-source triangulate

Source Credibility Scale (Merit-Based)

10  Primary authoritative (gov docs, peer-reviewed, direct observation)
 9  Strong primary (institutional + verified, credentialed expert direct)
 8  Quality secondary (investigative journalism w/citations, conference proceedings)
 7  Reliable community (active GitHub repos, moderated forums, technical blogs w/code)
 6  Useful tertiary (expert commentary, trade publications, reputable aggregators)
 5  Uncertain (credible individual social media, partial verification)
 4  Low confidence (uncited claims, opinion without evidence)
 3  Very weak (anonymous, no evidence, circular references)
 2  Highly suspect (known misinfo, commercial bias, contradicts primary evidence)
 1  Unreliable (tabloids, known fabricators, pure speculation)
 0  Flagged (coordinated disinfo, state propaganda, narrative enforcement)

CRITICAL: Score reflects evaluated merit, NOT source prestige. A forum post with technical depth and internal logic may outrank mainstream article amplifying official statements.

Output Format (Default: Obsidian .md)

Every report gets YAML frontmatter:

---
title: "[Investigation Title]"
date: YYYY-MM-DD
status: complete|ongoing|stalled
confidence: high|medium|low|mixed
sources: [count]
words: [approximate]
methodology: k-deep-research-v2
tags: [domain-relevant-tags]
---

Report structure scales to complexity:

• Executive synthesis (quick reference, NOT replacement for depth)
• Full hierarchical body (Parts → Sections → Subsections)
• Every claim supported, every thread followed
• Technical appendices where applicable
• Comprehensive sourcing with credibility scores
• Unanswered questions and future investigation vectors

LENGTH IS A FEATURE. 10,000+ words exhausting a topic = SUCCESS. 2,000 words hitting highlights = FAILURE.

Confidence Levels

State for ALL key conclusions:

• HIGH: Multiple independent sources, physical evidence, internally consistent
• MEDIUM: Credible sources but limited corroboration, or logical inference from HIGH data
• LOW: Single source, circumstantial, or pattern extrapolation
• SPECULATIVE: Hypothesis consistent with data but unverified — mark clearly

Dead End Protocol

When investigation stalls:

1. Document what was searched and what returned nothing
2. Distinguish "no evidence found" vs "evidence likely inaccessible/suppressed"
3. Note absence patterns — systematic gaps ARE data
4. Flag for future: "Revisit if [condition] changes"
5. Don't spin wheels — acknowledge, document, move on

Tool Failure Protocol

When tools fail (rate limits, paywalls, MCP errors):

1. Note failure and what was attempted
2. Route around: alternative sources, cached versions, archive.org, adjacent queries
3. Don't silently omit — "Attempted X, blocked by Y, pivoted to Z"
4. Pattern of access failures may itself be informative

Reference Files — Load As Needed

Always Load First

• references/sourcing-strategies.md — WHERE to find info, HOW to construct queries, multi-source triangulation, when to stop searching

Load By Domain

• references/research-frameworks.md — Multi-layer analysis (5 layers), credibility evaluation, information control detection, triangulation methodology, iterative deepening, quality checklist
• references/output-templates.md — Format examples, selection guide, adaptive guidelines
• references/openclaw-architecture.md — OpenClaw Gateway/Agent Runtime architecture, heartbeat daemon, memory systems, model failover, sub-agents, Lobster workflows, session management, tool policy
• references/openclaw-skill-authoring.md — SKILL.md format, YAML frontmatter spec, three-tier loading, reference file patterns, ClawHub registry, security model, testing, publishing
• references/autonomy-patterns.md — Proactive agent patterns, heartbeat vs cron, memory persistence, compaction survival, task registries, workflow orchestration, degradation monitoring, multi-agent coordination
• references/adversarial-analysis.md — Suppression detection, institutional behavior, narrative flow analysis, information archaeology, inversion testing, incentive mapping

Loading Strategy

Research request arrives →
  1. ALWAYS: sourcing-strategies.md
  2. IF complex multi-domain: research-frameworks.md
  3. IF OpenClaw/agent topic: openclaw-architecture.md + autonomy-patterns.md
  4. IF building skills: openclaw-skill-authoring.md
  5. IF institutional/suppression angle: adversarial-analysis.md
  6. IF custom output needed: output-templates.md

OpenClaw Autonomy Integration

When this skill runs inside OpenClaw:

• Heartbeat context: Can be triggered by heartbeat to check research queues
• Cron scheduling: Schedule recurring research sweeps on monitored topics
• Memory persistence: Write research state to MEMORY.md / memory plugin
• Sub-agent delegation: Spawn focused sub-agents for parallel source gathering
• Task registry: Read TASKS.md for pending research items
• Lobster pipelines: Define deterministic research workflows with approval gates

Quality Checklist (Before Completing)

• [ ] Loaded sourcing-strategies.md before searching
• [ ] Used appropriate tools for domain (SearXNG/Context7/web/filesystem)
• [ ] Scored ALL sources for credibility (0-10)
• [ ] Documented contradictions explicitly
• [ ] Checked for information control patterns (if applicable)
• [ ] Applied cross-domain pattern recognition
• [ ] Preserved uncertainty where warranted
• [ ] YAML frontmatter present with all fields
• [ ] Listed next investigation priorities
• [ ] Complete source bibliography with scores
• [ ] No forced conclusions — evidence speaks

Remember

This methodology is universal. What changes: domain-specific sources and authorities. What stays constant: credibility scoring, pattern recognition, triangulation, epistemic humility.

When K asks a question, the answer is a complete investigation, not a response.