Claude Code + Obsidian + スキルグラフでローカルリサーチエンジンを構築する方法

i built a research system for 4 clients in the past 6 months. one media company (well-known, can't name them), two consulting firms, and one independent analyst. combined, they cut their research costs by about 60%. one of them fired 3 junior researchers and replaced them with this system + one senior editor who reviews the output. no fancy tools. no $500/mo subscriptions. no 47 open chrome tabs. just a folder of .md files, one ai agent (claude), and a system that takes one research question and produces a multi-angle analysis that would normally take a team 2 weeks. this article is the full breakdown. every file, every link, every template. by the end you'll have a working research engine you can build in 1 hour. what is a research skill graph? most people use ai for research like this: open chatgpt, type "research X for me", get a surface-level summary that sounds like a wikipedia intro, then spend 3 hours verifying it and filling the gaps yourself. that's not research. that's just a google search with extra steps. the problem isn't the ai. it's that you're giving it zero structure. no methodology, no evaluation criteria, no framework for thinking about the topic from different angles. you're hiring a genius with amnesia every single time. a research skill graph fixes this. it's a folder of interconnected markdown files where each file is one "knowledge node", one piece of your research system's brain. inside each file you use [[wikilinks]] (double-bracket links like [[source-evaluation]] or [[contrarian]]) to reference other nodes. when you point claude at this folder and give it a research question, it doesn't just google stuff. it follows the links, reads your methodology files, applies your source evaluation criteria, then analyzes the topic through 6 completely different lenses before synthesizing everything. the difference: a single prompt gives you a summary. a skill graph gives you a research department. why 6 lenses instead of one big prompt? this is the core idea. instead of asking "research topic X", you force the ai to rethink the same question 6 times from fundamentally different angles: technical: what do the numbers actually say? look at data only economic: follow the money. who pays, who profits, what incentives drive behavior historical: what patterns repeat? what's been tried before? what context is everyone forgetting geopolitical: zoom out to the global chessboard. which countries, which power dynamics contrarian: what if the consensus is wrong? who benefits from the current narrative first principles: forget everything. rebuild from fundamental truths only each lens produces findings that often contradict the others. and that tension between lenses is where the real insight lives. when i ran "why are birth rates collapsing globally" through this system, the technical lens said "crisis: the math is brutal", while the contrarian lens said "japan has had low fertility for 50 years and hasn't collapsed." neither is wrong. the truth lives in the tension. the folder structure you're going to build: /research-skill-graph ├── index.md ├── research-log.md ├── methodology/ │ ├── research-frameworks.md │ ├── source-evaluation.md │ ├── synthesis-rules.md │ └── contradiction-protocol.md ├── lenses/ │ ├── technical.md │ ├── economic.md │ ├── historical.md │ ├── geopolitical.md │ ├── contrarian.md │ └── first-principles.md ├── projects/ │ └── (one subfolder per research topic) ├── sources/ │ └── source-template.md └── knowledge/ ├── concepts.md └── data-points.md 20 files. 6 folders. that's your entire research department. go create this structure right now. open a folder on your desktop or in obsidian, make the subfolders, create empty .md files with those names. then come back and we'll fill every single one. file 1: index.md (the command center) this is the single most important file. every research starts here. it's not a table of contents, it's a briefing document that tells the ai who you are, what system you're using, and exactly how to execute. # Research Skill Graph — Command Center ## 1. Mission Deep research system that takes ONE question and produces a multi-angle analysis no single Google search or ChatGPT prompt could ever match. Instead of 50 open tabs and scattered notes, this system forces structured thinking through 6 research lenses, each one rethinking the question from a fundamentally different angle. Research Question: [PASTE YOUR QUESTION HERE] Scope: [DEFINE BOUNDARIES — what's in, what's out] Time Horizon: [how far back and forward are we looking?] Output Goal: [what decision does this research inform?] ## Prior Research (optional — for compound mode) Check [[research-log]] for previous research that may connect to this question. Relevant prior projects: [link any related projects from research-log.md, or leave empty for clean slate] ## 2. Node Map Every node below is a knowledge file. Read the relevant ones before executing any task. The [[wikilinks]] are clickable — follow them. ### Methodology - [[research-frameworks]] — how to approach different types of questions. start here to pick the right research structure - [[source-evaluation]] — criteria for judging if a source is worth trusting. tier system from primary data to random blog posts - [[synthesis-rules]] — how to combine findings across lenses without losing nuance. the hardest part of research - [[contradiction-protocol]] — what to do when sources disagree. this is where the real insights hide ### Lenses (the core engine) - [[technical]] — how does it work mechanically? what do the numbers actually say? strip away narrative, look at data - [[economic]] — follow the money. who pays, who profits, what markets move, what incentives drive behavior - [[historical]] — what patterns repeat? what's been tried before? what context does everyone forget? - [[geopolitical]] — which countries, which power dynamics, which alliances and conflicts shape this? - [[contrarian]] — what if the consensus is wrong? who benefits from the current narrative? what's nobody saying? - [[first-principles]] — forget everything you think you know. rebuild from fundamental truths only ### Outputs - [[executive-summary]] — the final synthesis. 500 words max. what did we learn, what does it mean, what's still unknown - [[deep-dive]] — the full analysis organized by lens, with cross-references and contradictions highlighted - [[key-players]] — people, organizations, countries that matter most on this topic - [[open-questions]] — what we STILL don't know after research. often more valuable than what we found ### Knowledge Base - [[concepts]] — key terms, definitions, mental models relevant to the research - [[data-points]] — hard numbers, statistics, metrics collected during research. always with source attribution ### Sources - [[source-template]] — template for processing raw sources into structured notes ## 3. Execution Instructions When given a research question: 1. Read this file completely. understand the scope and goal 2. Read [[research-frameworks]] to pick the right approach for this type of question 3. Read [[source-evaluation]] so you know what counts as good evidence 4. For EACH lens in order (technical → economic → historical → geopolitical → contrarian → first-principles): a. Read the lens file for its specific angle and questions b. Research the topic THROUGH that lens only c. Record findings, sources, and confidence level d. Note any contradictions with previous lenses 5. Read [[contradiction-protocol]] — resolve or document disagreements between lenses 6. Read [[synthesis-rules]] — combine everything 7. Produce all 4 output files: [[executive-summary]], [[deep-dive]], [[key-players]], [[open-questions]] 8. Update [[concepts]] and [[data-points]] with everything learned CRITICAL RULE: each lens must RETHINK the question, not just add more information. the technical lens and the contrarian lens should feel like they were written by two different researchers who disagree with each other. that tension is where insight lives. notice: the node map gives context with every link. not just "[[technical]] — data" but "[[technical]], how does it work mechanically? what do the numbers actually say?" that extra context helps the agent make decisions without opening every file for every task. file 2: research-frameworks.md (pick your approach) # Research Frameworks How to approach different types of research questions. Read this BEFORE starting any research to pick the right structure. ## Framework Selection ### Type 1: "Is X true?" (Verification) - Start with [[technical]] lens to establish what the data actually says - Then [[contrarian]] to stress-test the claim - Then [[historical]] for precedent - Best for: fact-checking, debunking, validating assumptions - Example: "Is nuclear fusion viable by 2035?" ### Type 2: "Why is X happening?" (Causal Analysis) - Start with [[historical]] to trace the roots - Then [[economic]] to find incentive structures - Then [[technical]] for mechanism - Then [[geopolitical]] for systemic forces - Then [[contrarian]] to challenge your causal chain - Best for: understanding trends, explaining phenomena - Example: "Why are birth rates collapsing globally?" ### Type 3: "What happens if X?" (Scenario Planning) - Start with [[first-principles]] to establish base assumptions - Then [[technical]] for constraints - Then [[economic]] for incentives - Then [[geopolitical]] for power dynamics - Best for: forecasting, strategy, decision-making - Example: "What happens to Europe if birth rates stay below replacement for 30 years?" ### Type 4: "What should I do about X?" (Decision Support) - Start with [[executive-summary]] of existing knowledge - Then run all 6 lenses in parallel - Rank options by lens agreement (if 5/6 lenses point the same way, high confidence) - Best for: investment decisions, policy, strategy ## Research Depth Levels ### Level 1: Quick Scan (30 min) - 3 lenses maximum - Top 5 sources only - Goal: directional understanding, not certainty ### Level 2: Standard Research (2-3 hours) - All 6 lenses - 15-25 sources - Cross-reference findings between lenses - Goal: informed opinion backed by evidence ### Level 3: Deep Dive (1-2 days) - All 6 lenses with sub-questions - 50+ sources including primary data - Interview/expert source integration - Contradiction resolution required - Goal: publishable analysis, decision-grade intelligence ## Source Collection Strategy For each lens: 1. Start with the BEST single source (see [[source-evaluation]]) 2. Find the source that DISAGREES most with #1 3. Find primary data that lets you judge between them 4. Record everything in [[data-points]] with attribution file 3: source-evaluation.md (trust tiers) this is what stops your research from being garbage-in-garbage-out. a 5-tier system for evaluating every source before using it. # Source Evaluation How to judge if a source is worth trusting. Apply this to EVERY source before using it in analysis. ## Source Tier System ### Tier 1: Primary Data (highest trust) - Raw datasets (UN, World Bank, national statistics offices) - Peer-reviewed studies with methodology visible - Financial filings, government records - Direct measurements and observations - USE FOR: [[data-points]], hard claims, base assumptions ### Tier 2: Expert Analysis - Reports from domain-specific research institutions - Books by recognized authorities in the field - Long-form investigative journalism with cited sources - Conference papers and working papers - USE FOR: interpretation, causal claims, framework building ### Tier 3: Informed Commentary - Expert blog posts and newsletters - Quality podcasts with domain experts - Think tank reports (check funding sources) - Industry publications - USE FOR: angles you hadn't considered, hypothesis generation ### Tier 4: General Media - Major news outlets - Wikipedia (good for overview, never for final claims) - Popular science writing - USE FOR: initial orientation only. always verify upstream ### Tier 5: Social/Anecdotal (lowest trust) - Twitter threads, Reddit posts - Personal anecdotes - Viral content - USE FOR: signal detection only. "people are talking about X" ≠ "X is true" ## Red Flags (downgrade any source by 1 tier if present) - No cited sources or methodology - Author has financial incentive in the conclusion - Published by organization with known agenda on the topic - Cherry-picked time frames or geographies - Conflates correlation with causation - Uses emotional language instead of evidence ## Evaluation Checklist For every key claim in your research, ask: 1. What tier is this source? 2. Can I find the same claim in a Tier 1 or Tier 2 source? 3. Who funded this research or publication? 4. What would the author lose if they were wrong? 5. Is this the BEST available evidence, or just the first I found? See [[contradiction-protocol]] when two credible sources disagree. file 4: synthesis-rules.md (combine without flattening) this is the hardest part. most people skip synthesis and just stack facts. this file forces actual thinking. # Synthesis Rules How to combine findings across lenses without losing nuance. This is the hardest part of research — most people skip it and just stack facts. ## The Synthesis Process ### Step 1: Lens Summary After completing all 6 lenses, write a ONE paragraph summary per lens: - What is this lens's main finding? - Confidence level: high / medium / low - What surprised you from this angle? ### Step 2: Agreement Map Identify where lenses AGREE: - If 4+ lenses point the same direction → high confidence finding - If 3 lenses agree → moderate confidence, worth stating with caveats - If only 1-2 lenses support a claim → hypothesis only, flag as uncertain ### Step 3: Tension Map Identify where lenses DISAGREE: - [[technical]] says X but [[economic]] says Y → this tension IS the insight - Don't resolve by picking a winner. document both positions - Ask: "under what conditions is each lens correct?" - See [[contradiction-protocol]] for resolution framework ### Step 4: Second-Order Insights The best findings come from COMBINING lenses: - "The technical data shows declining fertility, but the economic lens reveals that financial incentives haven't reversed it anywhere — which means the [[first-principles]] lens's argument about cultural shifts might be the dominant factor" - These cross-lens insights are what make this system more powerful than reading 50 articles ### Step 5: Confidence Calibration For each major finding, state: - CLAIM: what you believe is true - EVIDENCE: strongest supporting data (from [[data-points]]) - CONFIDENCE: high / medium / low - WHAT WOULD CHANGE MY MIND: the specific evidence that would reverse this conclusion ## Output Rules - Never present a single-lens finding as a conclusion - Always show the tension between lenses - Separate "what the data shows" from "what I interpret" - [[open-questions]] is as important as [[executive-summary]] - Prefer "it seems likely that X because Y and Z, though A complicates this" over "X is true" ## Anti-Patterns (things that kill good research) - Confirmation bias: only searching for evidence that supports your initial hunch - Narrative fallacy: making a clean story out of messy, contradictory data - Recency bias: overweighting the latest article over 30 years of data - Authority bias: believing something because an impressive person said it - Anchoring: letting the first number you found define your mental range file 5: contradiction-protocol.md (where the real insights hide) most research buries contradictions. this system surfaces them on purpose. contradictions are features. # Contradiction Protocol What to do when sources or lenses disagree. This is where the real insights hide — contradictions are features, not bugs. ## When Two Sources Disagree ### Step 1: Check the basics - Are they talking about the same thing? (different geographies, time periods, definitions) - Are they using the same data? (same dataset, different interpretation?) - Is one source a higher tier than the other? (see [[source-evaluation]]) ### Step 2: Find the root of disagreement Usually one of: - Different data: they're looking at different datasets → find out which is more complete/recent - Different interpretation: same data, different conclusions → examine their reasoning chain - Different scope: one is global, one is country-specific → both might be right in their context - Different timeframe: short-term vs long-term trends can look opposite - Different incentives: one has reason to spin the data → check funding, affiliations ### Step 3: Document, don't resolve In [[deep-dive]], write it as: "Source A argues [X] based on [data]. Source B argues [Y] based on [data]. The disagreement likely stems from [root cause]. Under conditions [C1], A is probably right. Under conditions [C2], B is probably right." ### Step 4: Upgrade to [[open-questions]] if unresolvable If after honest effort you can't determine who's right: - This IS a finding. "We don't know whether X or Y" is valuable intelligence - Add it to [[open-questions]] with the specific evidence needed to resolve it - This often becomes the most interesting part of the research ## When Two Lenses Disagree This is expected and healthy. Different lenses SHOULD produce tension. ## Confidence Adjustment - 2 lenses agree, 1 disagrees → investigate the disagreeing lens deeper. it might see something the others miss - All lenses agree → be suspicious. you might have confirmation bias. re-read [[contrarian]] - No lenses agree → your research question might be too broad. narrow the scope in [[index]] files 6-11: the 6 lenses every lens file follows the same structure: core questions, how to research through that angle, output format, voice, and connections to other lenses. i'll show the technical lens as the detailed example, the other 5 follow the exact same pattern, you just swap the angle. # Lens: Technical Strip away opinions and narratives. What do the numbers actually say? What mechanisms are at work? ## Core Questions 1. What does the DATA show? (not what people say about the data) 2. What are the measurable inputs and outputs? 3. What mechanisms drive the phenomenon? 4. What are the hard constraints (physical, biological, mathematical)? 5. What metrics matter most and how are they measured? 6. Where is the data incomplete or poorly measured? ## How to Research Through This Lens - Start with Tier 1 sources ONLY (see [[source-evaluation]]): raw datasets, peer-reviewed studies - Look for: time series, geographic comparisons, demographic breakdowns - Quantify everything. replace "declining" with "declined by X% between Y and Z" - Identify measurement problems: how is this data collected? what's excluded? - Find the base rates. what's "normal"? what's the historical range? ## Output Format For each finding: - METRIC: [what you measured] - VALUE: [the number] - SOURCE: [Tier 1 or 2 source with date] - TREND: [direction and rate of change] - CAVEAT: [measurement limitations] Record all hard numbers in [[data-points]]. ## Voice Clinical. precise. no emotional language. "fertility rate dropped from 2.1 to 1.6" not "fertility is collapsing." let the numbers speak. ## Connects To - [[source-evaluation]] — only high-tier sources for this lens - [[data-points]] — all numbers go here - [[economic]] — technical data often explains economic outcomes - [[first-principles]] — technical constraints define what's possible for the other 5 lenses, same structure but different core questions: economic: who pays, who profits, what incentives, what policies tried, what ROI? historical: when has this happened before, what conditions, what was tried, what worked? geopolitical: which countries most affected, what power shifts, what alliances change? contrarian: what's the mainstream narrative, what's the strongest argument against it, who benefits from the consensus? first principles: what are the absolute base-level facts, what's the simplest model that explains 80% of what we observe? file 12: source-template.md copy this for each major source you process during research. # Source Template ## Source: [Title] - Author: [who] - Date: [when published] - URL: [link] - Tier: [1-5, from [[source-evaluation]]] - Lens: [which lens found this source] ## Key Claims 1. [claim with page/section reference] 2. [claim] 3. [claim] ## Data Extracted [any numbers → also add to [[data-points]]] ## Methodology Notes [how did they arrive at their conclusions? any red flags?] ## Connections [what other sources or concepts does this relate to? use [[wikilinks]]] ## My Assessment [is this trustworthy? what's strong, what's weak?] the compound effect (why this gets better over time) this is what makes the system fundamentally different from chatgpt conversations or google research. knowledge/concepts.md and knowledge/data-points.md accumulate across ALL your research projects. after 5 projects, your ai starts with a base of 200+ verified data points and 50+ defined concepts. research-log.md tracks every project with key findings and connections. your 10th project doesn't start from zero, it starts from everything you've already learned. even better: the open-questions from one research become the index.md of the next. when i researched birth rates, one open question was "will ai automation offset labor shortages fast enough?", that's an entire research project that already has context from the demographic data. and if you want a clean slate? just upload only methodology/ and lenses/ to a fresh claude project. no knowledge/, no research-log. same system, fresh brain. how to use it method 1: claude projects (recommended). create a new project, upload all files, give it a topic with "follow the execution instructions in index.md." method 2: paste context. copy index.md + relevant lens files into any ai chat. less powerful but works anywhere. method 3: claude code + obsidian (most powerful). point claude code at your local vault. the agent reads and writes files directly. the graph evolves itself, updating knowledge files, adding new concepts, refining lenses based on output quality. fully autonomous. visualize with obsidian download obsidian (free) from obsidian.md. open your research-skill-graph folder as a vault. you immediately see a graph view showing how all your nodes connect. index.md sits at the center. the 6 lenses radiate out. methodology files connect to lenses. knowledge files connect to everything. two things this gives you: you can see which nodes are disconnected (research gaps) and spot unexpected connections between projects (compound insights). obsidian is optional, the ai reads markdown regardless. but the visual graph makes the system tangible and helps you debug it. why this beats traditional research traditional research: open 50 tabs, read 20 articles that all say the same thing, miss the contrarian view, miss the historical pattern, get lost in confirmation bias, produce a summary that sounds smart but has no structure. skill graph research: one question goes through 6 forced angles, each angle has evaluation criteria, contradictions are documented not hidden, sources are tiered, findings compound across projects. the biggest difference is the contrarian lens. traditional research almost never challenges its own findings. this system has a built-in devil's advocate that asks "what if everything i just found is wrong?", and rates the strength of the counter-argument honestly. build it in 1 hour: create the folder structure. 20 files, 6 folders. takes 5 minutes fill index.md first: this defines everything else fill the 6 lens files with the core questions and output formats fill methodology/ with source evaluation tiers and synthesis rules upload everything to a claude project give it a topic and test iterate. update the lens files based on output quality. add concepts to knowledge/ after each project the system gets better every time you use it. that's the whole point.