3.7 · Research & Analysis ở scale lớn

Mở đầu: "Tôi không có thời gian đọc hết đống này"

• 40 PDF paper research chưa đọc
• 50 interview transcripts team user research thu thập
• Data file 2 triệu rows của usage metrics
• 100 contracts cần cross-check compliance
• Year worth of meeting notes — "tôi quên decisions nào đã ra rồi"

Trong chat

Paste 40 PDF? Impossible. Upload limit, context limit. Chat "ngộp".

Cowork

"Read everything in /papers/, find patterns, flag contradictions, output synthesis."

Task 2 ngày → 3 giờ. Đây là chỗ Cowork shine where chat can't as easily.

3 shapes of work Cowork handle naturally

Chat và Cowork đều có thể giúp research/analysis. Khác biệt ở shape of work fit mỗi cái.

Chat phù hợp khi:

• Material fit trong 1 conversation
• Bạn paste được, discuss được, iterate turn-by-turn
• Work là conversational refinement

Cowork fit hơn khi có 1 hoặc nhiều những shape này:

┌────────────────────────────────────────────────────────┐
│                                                        │
│   1. 📦 VOLUME                                          │
│      Too many files, hoặc files too large              │
│      Cowork reads in place, không bị bound bởi         │
│      single chat window.                               │
│                                                        │
│   2. 🔀 PARALLELISM                                     │
│      Task = "do same analysis across N items"          │
│      Cowork processes simultaneously, không sequential │
│                                                        │
│   3. 💻 IN-PLACE COMPUTATION                            │
│      Run code on files where they live                 │
│      Write results back to disk — no upload/download   │
│                                                        │
└────────────────────────────────────────────────────────┘

Cowork không "làm được điều chat không thể"

Đây là nuance quan trọng:

None of these are things chat can't touch — they're things Cowork is built to handle more naturally when they're the shape of the work.

Chat có thể tóm tắt 5 papers (paste từng cái). Cowork tóm tắt 50 papers (read from folder). Cùng capability, different fit for shape.

Shape #1: Volume — Quá nhiều để paste

Scenarios

• Research synthesis: 40 academic papers về 1 topic
• Interview analysis: 50 user interview transcripts
• Literature review: 100+ sources cho grant application
• Market intelligence: Hàng chục reports từ consulting firms
• Log analysis: Weeks of application logs

Chat vs Cowork quantitative

Ví dụ prompt cho volume

Read everything in the research folder and write a synthesis report. Where do sources agree? Where do they contradict each other? Flag any claim that only one source makes. Output: synthesis.docx with citations [Author, Year].

Cowork approach:

• Reads 40 papers (10-30 phút tùy số lượng + complexity)
• Holds all in "working memory" via subagent architecture
• Finds cross-references a human would miss scanning sequentially
• Outputs coherent narrative with bibliography

Shape #2: Parallelism — N items cùng analysis

Scenarios

• Compare 4 vendors across 5 dimensions
• Analyze 10 competitor companies
• Review 50 contracts against 1 template
• Process 30 candidate resumes against JD
• Run same analysis on 20 portfolio companies

Subagents under the hood

Bạn đã gặp subagents sơ bộ ở **Bài 3.2. Giờ đi sâu: khi task có independent pieces, Cowork spawn subagents. Mỗi cái fresh context, focused attention**:

Task: "Compare 4 vendors on pricing, integrations, reviews"

          ┌─────────────────────────┐
          │    Main orchestrator    │
          └────────────┬────────────┘
                       │
        ┌──────────────┼──────────────┬────────────┐
        ▼              ▼              ▼            ▼
    ┌───────┐      ┌───────┐      ┌───────┐    ┌───────┐
    │ Agent │      │ Agent │      │ Agent │    │ Agent │
    │   A   │      │   B   │      │   C   │    │   D   │
    └───┬───┘      └───┬───┘      └───┬───┘    └───┬───┘
        │              │              │            │
  Pricing A      Pricing B       Pricing C    Pricing D
  Integr. A      Integr. B       Integr. C    Integr. D
  Reviews A      Reviews B       Reviews C    Reviews D
        │              │              │            │
        └──────────────┴──────────────┴────────────┘
                       │
                       ▼
              ┌────────────────┐
              │  Synthesis     │
              │  + Final output│
              └────────────────┘

Lợi ích quan trọng:

Tasks too large to hold in one conversation can be split into pieces that each fit comfortably, and each piece gets focused attention.

Ví dụ cụ thể: Vendor D analysis isn't diluted by Vendor A's details. Mỗi subagent focus độc lập.

Khi nào bạn "get" subagents tự động?

Cowork auto-deploy subagents khi:

• Prompt structure rõ là "do X for each Y" pattern
• Folder chứa similar files (all PDFs, all transcripts, all reports)
• User explicitly ask parallel ("process each file simultaneously")

Prompt to trigger parallel work

Implicit (Cowork decides):

Compare vendors A, B, C, D trên pricing, integrations, reviews.

Explicit (make sure parallel):

"For each vendor in the folder, research: pricing, integrations, recent reviews. Process in parallel — không sequential. Synthesize findings into comparison matrix + recommendation memo."

Shape #3: In-place computation

Scenario

Data lives trên drive của bạn. Bạn cần compute on it — statistical analysis, transformations, aggregations, charts.

Chat: Upload file → Claude respond → Copy text xuống → Paste vào Excel → Chart manually.

Cowork: "Compute X on this file, save results back to disk."

Power unlock

• No upload/download loop — file ngay trên disk
• Run code (Python, pandas, matplotlib...) trong isolated environment
• Write results back to folder ngay
• Charts, pivots, formulas inserted vào .xlsx native

Ví dụ: Customer churn analysis

Prompt:

Trong file usage-data.csv (2M rows): - Calculate month-over-month usage change per customer_id - Flag accounts với 3 tháng liên tiếp decline (at-risk) - Create scatter plot: usage trend vs contract_value - Cluster at-risk accounts bằng k-means (k=5) để find patterns Output: - at-risk-accounts.xlsx (list + metadata) - churn-analysis.docx (findings + charts) - scatter-plot.png

Cowork:

1. Read CSV
2. Pandas manipulate
3. matplotlib chart
4. scikit-learn cluster
5. Export multiple formats
6. Output all in folder

Thời gian: 3-4 giờ data scientist manual → 15 phút Cowork + review.

4 loại research/analysis tasks

Taxonomy giúp bạn recognize which shape fits:

1. Research Synthesis

Definition

Combine notes, articles, papers, saved research → coherent report.

Value at scale

Cross-referencing. Cowork holds all sources at once, finds connections bạn miss khi đọc sequentially.

Prompt pattern:

Read everything trong [research folder].
Write synthesis report:
- Where sources agree
- Where they contradict
- Claims made by only 1 source (flag these)

Output: synthesis.docx with citations.

Use cases:

• Literature review cho grant application (80 papers → synthesis)
• Industry trend report từ saved articles
• Competitor intelligence aggregation
• Academic research paper drafting

2. Transcript Analysis

Definition: Extract themes, decisions, action items từ meeting notes, interviews, recordings.

Value at scale:

• 5 transcripts → chat OK
• 50 transcripts → Cowork reads parallel, finds places they disagree
• That disagreement = insight

Prompt pattern:

Read all transcripts trong [folder].

Tell me:
- What did MOST people agree on?
- Who disagreed?
- What do disagreers have in common? (demographic, role, use case...)

Output: transcript-themes.docx with quotes as evidence.

Key insight:

You might learn that the outliers share a trait. That's the insight buried at scale.

Use cases:

• 50 user interviews → find personas naturally grouped
• 30 sales call transcripts → objection patterns
• 20 customer support calls → issue clustering
• 10 board meetings → decision drift over time

3. Data Analysis

Definition

Statistical work trên data files — outliers, cross-tabulation, time-series, clustering.

Value at scale

In-place computation.

Prompt pattern:

For [each X] trong [data file]:
- Calculate [metric]
- Flag [condition]
- Compare against [benchmark]

Output:
- [data table] với flagged rows
- [report] với findings
- [visualization] nếu applicable

Use cases:

• Customer health scoring across book of accounts
• Product usage anomaly detection
• Financial variance analysis
• A/B test results interpretation
• Cohort analysis

4. Knowledge Synthesis (Your Own Work)

Definition

Point Cowork tại your accumulated notes, journals, project files. Ask what patterns it finds.

Meta-angle

You don't always know what's in your own files. Cowork surfaces patterns.

Prompt pattern:

Read tất cả my notes trong [notes folder] từ [timeframe].

Find:
- Questions tôi keep asking
- Decisions contradicting each other
- Open loops never resolved
- Themes emerging across projects

Output: self-reflection-report.md

Use cases:

• Quarter reflection: "Where did my time go?"
• Decision audit: "Have I been consistent?"
• Learning synthesis: "What's the common thread in what I'm studying?"
• Journal mining: insights từ personal journaling

Shift tư duy: Ask for Signal, Not Summary

Quan trọng nhất của bài này. Prompt hữu ích nhất không phải "summarize everything".

Là câu hỏi bạn sẽ answer nếu có thời gian.

Ví dụ sharpening

┌────────────────────────────────────────────────────────┐
│                                                        │
│  😐 VAGUE (waste of Cowork)                             │
│                                                        │
│  "Summarize these transcripts"                         │
│                                                        │
│           ↓ SHARPEN ↓                                  │
│                                                        │
│  🎯 SIGNAL-SEEKING                                      │
│                                                        │
│  "What did most people agree on?                       │
│   Who disagreed? What do they have in common?"         │
│                                                        │
└────────────────────────────────────────────────────────┘

More sharpening examples

Test: Prompt của bạn có một câu trả lời actionable không? Nếu no → sharpen more.

Framework: 4 sharp question types

┌────────────────────────────────────────────────────┐
│                                                    │
│   1. 🔀 CONTRADICTION                               │
│      "What contradicts what?"                      │
│                                                    │
│   2. 🎯 OUTLIER                                     │
│      "What's different? Who's the exception?"      │
│                                                    │
│   3. 🔮 PREDICTION                                  │
│      "What would change my decision?"              │
│                                                    │
│   4. 📈 PATTERN                                     │
│      "What pattern keeps showing up?"              │
│                                                    │
└────────────────────────────────────────────────────┘

Use these as prompt starters cho research tasks.

Walkthrough đầy đủ: 40 papers → synthesis report

Setup

Folder /research-genai-productivity/:

• 40 academic papers (PDF)
• meta.md — your notes on what's in folder
• CLAUDE.md — project instructions

Prompt

Literature review synthesis task. Folder: /research-genai-productivity/ (40 papers, 2023-2025) Phase 1: Catalog mỗi paper với: author, year, methodology, key claims, sample size. Phase 2: Identify themes — group papers theo sub-topic. Expected sub-topics (but don't limit to): knowledge worker productivity gains, AI literacy correlation, cognitive offloading risks, team-level impacts. Phase 3: Synthesis narrative: - Area of consensus (claim X with evidence count) - Area of contested (Y with for/against) - Research gaps (what no one answered) - Methodological concerns (studies với weak design) Citations: [Author, Year] inline. Full bibliography at end. Output: synthesis-report.docx (3000-4000 words).

What Cowork does

Phase 1 (10 min): Parallel subagents scan each PDF → extract metadata
                  Output: paper-catalog.xlsx (40 rows)

Phase 2 (8 min):  Cluster themes across catalog
                  Output: theme-groups.md

Phase 3 (15 min): Synthesis
                  Output: synthesis-report.docx

Output structure

synthesis-report.docx (3,200 words)

1. Executive Summary (200 words)
2. Area of Consensus (800 words)
   - 12 papers claim "GenAI gives 20-40% productivity bump..."
     [citations]
   - 8 papers converge on "AI literacy matters more than raw 
     access..." [citations]
   
3. Contested Areas (1000 words)
   - "Junior vs Senior impact" — 5 papers argue junior gains 
     most [citations], 4 argue senior gains most [citations]
   
4. Research Gaps (500 words)
   - "No longitudinal study >12 months"
   - "Enterprise deployment data behind paywalls"
   
5. Methodology Concerns (400 words)
   - Lab studies dominate → ecological validity question
   
6. Bibliography (300 words, 40 entries)

Thời gian

• Manual: 2 tuần
• Cowork: 33 min Cowork + 2 giờ review = ~2.5 giờ total

Prompt templates — Research & Analysis

Template 1: Synthesis với compare/contrast

Read all [file type] trong [folder].

Analyze:
- Where do sources agree? (with evidence count)
- Where do they contradict? (with specific claim quotes)
- What claim is made by only 1 source? (flag for verification)

Output: synthesis.docx với citations [Author, Year].
Length: [X words]

Template 2: Theme extraction from transcripts

Read all transcripts trong [folder].

Identify:
- 3-5 dominant themes (with supporting quote from each source)
- Outliers: which transcripts don't fit themes?
- What trait do outliers share?

Output: themes-report.md
Format: theme → evidence → implications

Template 3: Cross-doc search for inconsistency

Check tất cả [document type] trong [folder]:

- Tên feature/term used inconsistently?
- Version info outdated?
- Claims tô hồng/contradicting across docs?
- Links broken?

Output: consistency-report.md với findings + suggested fixes.
Priority rating: High/Medium/Low cho mỗi issue.

Template 4: Data pattern detection

Analyze [data file]:

Find:
- Clusters in [dimension 1] (visualize)
- Trends over [time dimension] (chart)
- Outliers in [metric] (flag)

For each pattern:
- Claim
- Evidence (data subset)
- Implication

Output: report.md + charts folder.

Template 5: Benchmark comparison

Benchmark [subject] against:
- [competitor 1]
- [competitor 2]
- [competitor 3]

Dimensions:
- [Dim 1] (metric: X)
- [Dim 2] (metric: Y)
- [Dim 3] (metric: Z)

Parallel process each competitor (separate subagent).

Output: 
- comparison-matrix.xlsx
- competitive-summary.docx
- Recommendation với confidence level

Template 6: Market intel scan

Scan market intelligence [industry/segment]:

1. Product launches từ top N players (last 90 days)
2. Funding/M&A activity
3. Regulatory changes
4. Tech/platform shifts

Sources: industry news sites, press releases,
SEC filings (if applicable), web search.

Output:
- intel-briefing.docx (3-5 pages)
- tracker.xlsx (data + sources)

Template 7: Investment/portfolio analysis

Cover [N] companies trong [sector]:

For each (parallel):
- Read 10-K / annual report
- Earnings calls latest quarter
- Analyst reports
- Recent news

Synthesize:
- Cross-company trends
- Winners/losers
- Red flags
- Investment memo với top picks

Output:
- individual-memos/ (one per company)
- sector-synthesis.docx
- watchlist.xlsx với metrics

Template 8: Customer voice analysis

Read all [customer inputs]:
- Support tickets (tool export)
- NPS comments (file)
- App store reviews (scraped)

Extract:
- Top 10 recurring complaints (rank by frequency)
- Top 10 feature requests (rank by volume)
- Sentiment by segment (chart)
- Quote exemplars for each (representative)

Output:
- customer-voice-report.docx
- ticket-cluster.xlsx
- quote-library.md

Template 9: Audit / compliance scan

Audit [N documents] against [standards/regulations].

For each document:
- Compliance score (0-100)
- Gaps identified
- Remediation needed

Output:
- compliance-matrix.xlsx (all docs × all criteria)
- High-risk findings.docx (top 10 critical)
- Action plan.md với priorities

Template 10: Self-knowledge synthesis

Read my notes/journals/project files trong [folder] 
từ [timeframe].

Find:
- Questions I keep asking but never answer
- Decisions I made trong project A contradicting project B
- Open loops never closed
- Themes emerging across contexts

Output: self-reflection.md với sensitive/private flags.

Case studies đa dạng

Academic Researcher — Grant literature review

Challenge

80 papers cần cover cho literature review grant application.

Prompt

Template 1 (Synthesis) + Phase structure.

Approach:

• Phase 1: Catalog 80 papers (30 min)
• Phase 2: Theme grouping (20 min)
• Phase 3: Write 8000-word literature review với citations (45 min)

Kết quả: 3 tuần → 1 tuần (1.5 ngày Cowork + 5 ngày human polish).

Investment Analyst — Sector coverage

Challenge

15 portfolio companies cần quarterly deep review.

Prompt

Template 7 (parallel per company).

Approach:

• 15 subagents, 1/company
• Each: read 10-K + earnings call + analyst reports
• Synthesize cross-company trends

Kết quả: 1 tuần → 2 ngày.

Clinical Research Coordinator — Patient record consolidation

Challenge: 200 patient folders, messy organization.

Prompt:

For each patient folder in /patients/:
- Pull lab results across all tests
- Extract imaging notes
- Identify current medications
- Flag abnormal values

Timeline-based output per patient.
Summary dashboard across all patients.

Kết quả: 30 phút/patient → 5 phút review. 200 patients: 100 giờ → 16 giờ.

Product Manager — User interview synthesis

Challenge

50 user interviews về new onboarding flow.

Prompt

Template 2 + custom:

For 50 transcripts trong /user-interviews/:

- 3-5 dominant themes about onboarding pain
- Segment by user tier (free/pro/enterprise)
- Find outliers: users who DIDN'T complain (what's different?)
- Extract actionable feature requests (rank by volume)

Output: user-research-deck.pptx (15 slides cho product review meeting).

Kết quả: 3 ngày read/tag → 3 giờ.

Compliance Officer — Quarterly policy audit

Challenge

50 policies cần compliance review quarterly.

Prompt

Template 9 (audit) parallel.

Approach:

• 50 subagents (one/policy)
• Check against current reg standards
• Score adherence

Kết quả: 2 tuần → 3 ngày.

Customer Experience Analyst — Multi-source review

Challenge

Reviews scattered across Amazon, Shopify, social, support tickets.

Prompt

Template 8 (customer voice).

Approach:

• Aggregate from all sources
• Sentiment analysis
• Theme extraction
• Weekly digest for product team

Kết quả: 2 ngày manual → 4 giờ review.

Growth Marketer — Multi-platform performance

Challenge: 5 ad platforms (Meta, Google, LinkedIn, TikTok, X) — weekly analysis.

Prompt:

Pull data from past 7 days ad platforms.
For each platform:
- Spend vs CPA vs CTR
- Top 3 performing ads
- Worst 3 (pause candidates)
- Anomalies (spend spike, CTR drop)

Cross-platform synthesis:
- Best performing creative theme?
- Best performing audience segment?
- Budget reallocation recommendation

Output: weekly-ad-report.xlsx + executive-summary.md.

Kết quả: 1 ngày → 1 giờ review.

Anti-patterns khi research/analysis

Prompt ask for "summary" thay vì "signal"

Triệu chứng

"Summarize the research folder".

Tại sao sai

Summary = compressed list of what each source said. Không actionable. Không thấy signal.

Cách đúng

Dùng sharpening framework. Ask for contradiction, outlier, prediction, pattern.

Không tận dụng parallelism

Triệu chứng

"Analyze each paper one by one in order".

Tại sao sai

Sequential = slow. Cowork có subagent capability không dùng lãng phí.

Cách đúng

Make parallel explicit: "for each ... in parallel" or design task shape để auto-parallel.

Expecting chat-like iteration on huge corpus

Triệu chứng

Start task, nhận 30% output, chat back and forth nhiều times refine.

Tại sao sai

Khi Cowork đã đọc 40 papers, re-run = tốn time + allocation.

Cách đúng

Front-load specificity trong prompt. Clear output format, structure, constraints. Refine ở output-level (edit document), không Cowork-level (redo task).

Không provide meta về folder contents

Triệu chứng

Folder có 100 files mixed quality + relevance, prompt just "read everything".

Tại sao sai

Cowork đọc cả noise. Output dilute.

Cách đúng

Maintain meta.md trong folder: "These 100 files are [type], curated for [purpose]. Priority: A-papers (peer-reviewed), B-papers (preprints), C-papers (blog posts)".

Không verify outliers/contradictions

Triệu chứng

Cowork flag "Source X contradicts Source Y" → bạn accept không verify.

Tại sao sai

Even at scale, Cowork can miss nuance. Contradiction có thể là Cowork's interpretation không exact.

Cách đúng

Output có citations → sample 3-5 random contradictions, verify từ source file. Build trust incrementally.

Task scope quá rộng

Triệu chứng

"Read everything I've ever written and tell me what to do next."

Tại sao sai

No bound = low quality. Cowork can't surface useful signal without scope.

Cách đúng

Narrow: "Read my Q1 2026 journal entries. Find 3 specific projects I mentioned but haven't advanced. Prioritize."

Mẹo nâng cao

Mẹo 1: Multi-phase prompt cho deep synthesis

Phase 1: Catalog (extract metadata per source)
Phase 2: Cluster (group by theme)
Phase 3: Analyze (depth on each theme)
Phase 4: Synthesize (narrative across themes)

Output each phase separately — I'll review before next phase.

Quality compounds. Each phase validates before next.

Mẹo 2: Specify "confidence levels" in output

For each claim in your synthesis, rate confidence: - HIGH: multiple primary sources converge - MEDIUM: single strong source or multiple weak - LOW: inference from limited evidence Mark với emoji trong output.

Reader can weight insights appropriately.

Mẹo 3: Ask for "what would change your mind"

For top 3 conclusions, specify: what evidence would change them? What if the next 10 papers show opposite?

Surfaces epistemic humility. Forces Cowork beyond surface claims.

Mẹo 4: Trigger disagreement intentionally

After synthesis, play devil's advocate. What's the strongest counter-argument to your main conclusion?

Single prompt → 2 perspectives. Richer thinking.

Mẹo 5: Save synthesis + rationale for audit

Output include: evidence-log.md listing, for each major claim, which source(s) it came from

so I can audit trail."

Critical cho legal/regulatory/academic work.

Mẹo 6: Segment analysis khi dataset heterogeneous

Don't synthesize 40 papers where 20 are lab studies and 20 are industry surveys. Sep:

Segment sources theo methodology. Synthesize WITHIN segment. Then cross-segment compare.

Mẹo 7: Use visualization cho multi-dimensional data

Output: scatter plot (x=metric A, y=metric B, color=segment). Find clusters. Annotate outliers.

Cowork generate real PNG. Human brain processes viz faster than tables.

Mẹo 8: Package research workflows thành skills

Nếu bạn làm research synthesis recurrently, convert pattern thành skill (học ở **Bài 3.4**). Save prompt construction mỗi lần — /research-synthesis [folder] đủ invoke.

Áp dụng ngay

Bài tập 1: Sharpen a vague prompt (~5 phút)

Take 1 recent vague ask bạn đã làm cho Claude. Apply sharpening:

Vague version: _______________________

Sharpened version (with contradiction/outlier/prediction/pattern framing): _______________________

Bài tập 2: Run a research task (~20 phút)

Identify body of input bạn đã putting off working through:

• Folder of research papers / reports
• Collection of interview transcripts
• Historical meeting notes
• Saved research bookmarks

Prompt with sharp question, not summary:

[Your specific signal-seeking question]. Process in parallel where possible. Output: [format] với citations. Length: [X words].

Run. Review:

• Was output actionable? (not just descriptive)
• Were outliers/contradictions flagged?
• Would you make different decision based on this?

Bài tập 3 (optional): Build research skill (~15 phút)

If you do research synthesis recurrently:

1. Identify your "standard shape" of research task
2. Convert into skill:

/create-skill "research-synthesis — for folder of papers,
synthesize with consensus/contested/gaps framework,
output docx with citations, include confidence levels"

1. Next research task: /research-synthesis [folder]

Tóm tắt bài học

Tài liệu tham khảo

• Subagents architecture deep dive — How parallel work happens
• AI Fluency course — Discernment & research habits
• Cowork use-cases: research