In spring 2023, a startup CEO in Seoul's Gangnam district was transferring prices from competitor websites into a spreadsheet until two in the morning.

With about thirty tabs open, he read App Store reviews, then moved to Blind's forum to scan job postings, then checked investment history on Crunchbase. He repeated this work every month.

In offices in other cities, similar scenes were unfolding at similar times. An investment analyst was preparing morning briefings by moving through dozens of news sites, a graduate student was filtering hundreds of papers on PubMed while drafting a literature review, and a marketing director was sifting through past campaign data to estimate the success probability of the next project.

These four tasks share a common structure.

Gathering vast amounts of information, filtering out noise, identifying patterns, and organizing the result into a form that humans can judge. They take a long time because information is scattered, and mistakes happen because human attention has limits.

Cowork handles collection and organization in these four research tasks.

It opens browsers to visit competitor websites, reads news feeds to create summaries, analyzes papers to create comparison tables, and builds predictive models from historical data. What remains for humans is to question the machine's organized results, interpret them, and make decisions.

Information gathering and intelligence are different. Information gathering is collecting "what happened," while intelligence is judging "what does this mean for us."

Cowork handles the former at overwhelming speed. The latter remains human territory. When reading a report organized by a machine, you need an eye to observe what is absent from the text as much as what is written.

1 Automated Competitive Analysis

A Why It Is Needed

Competitive analysis is not about capturing and pasting a few screenshots. It is closer to building a reconnaissance system.

You must find out whether competitors changed prices, launched new features, hired for certain roles, and where customers express complaints. This information is not in one place. It is scattered across websites, blogs, social media, app stores, review platforms, and job postings.

Two problems arise when humans do this work. One is time. To analyze three competitors, you must check each company's pricing page, compare feature lists, read customer ratings on review sites like G2 or Capterra, and track hiring trends on LinkedIn.

Transferring this data to Excel to create a comparison table and organizing it into a report executives can read takes at least several days. By the time the report is finished and lands on the desk, the market has already moved one step further.

The other is consistency. When tired, humans skip a tab. They focus on familiar competitors and miss new entrants. An analysis written Friday afternoon differs in depth from one written Monday morning. Cowork does not skip. It investigates every item on the list with equal depth.

B What It Does

Cowork's competitive analysis combines two technologies. One is Claude in Chrome, a browser extension that allows AI to directly control Chrome. Cowork navigates to specified URLs, scrolls pages, reads price tables, and takes screenshots to analyze layouts. The other is parallel sub-agent processing. If there are three competitors, three sub-agents move simultaneously, and the main agent synthesizes the results.

There are two final outputs. An Excel matrix comparing prices, features, and target customers across competitors at a glance, and a Word-format intelligence report that includes strategic implications.

[Note] Legal Boundaries of Web Scraping

It is technically possible to collect public information from competitor websites using browser automation. However, "technical possibility" and "legal permission" are different matters. The Robots Exclusion Standard (robots.txt) has been used for over thirty years as a means for website owners to control automated access. You must first check the competitor's terms of service, robot policy, and access frequency limits. If an official API exists, prioritize it over HTML scraping. Exclude pages requiring login, paid content, and areas containing personal information from collection. Respecting the boundary between what is technically possible and what is legally permitted is the first principle of automation.

C How to Do It: Basic Use

Start by collecting public information from one competitor and creating a summary document.

Walkthrough

① Create a folder called "Competitor_Analysis" on your desktop.

② Open the Claude Desktop app and click the "Cowork" tab at the top.

③ In the left sidebar, grant read/write permissions to the "Competitor_Analysis" folder.

④ Check that Claude in Chrome is activated in Connector settings. The Chrome extension must be connected in Settings > Connectors.

⑤ Enter the following prompt.

"Visit competitor A's website (www.competitor-a.com) and collect the following information. 1) Prices by plan and included features from the pricing page, 2) Core value proposition and target customer message from the product overview page, 3) Product update related content from blog posts within the last three months. Organize the collected information into a Word file and save it to this folder."

⑥ Cowork creates a plan and displays it in the sidebar. Confirm the order "Visit pricing page → Visit product overview page → Explore blog" and approve execution.

⑦ The browser opens automatically and Cowork traverses each page. When a tab border turns orange, it means Cowork is controlling that tab. ⑧ When the work is complete, a Word document appears in the folder.

D How to Do It: Applied Examples

Expand to multiple competitors and add review data.

① Enter the following prompt.

"Analyze three competitors in the AI app builder market. Targets: Replit, Lovable, Bolt. Visit each company's website and collect the following.

1) Prices by plan and free trial conditions, 2) Core features and limitations, 3) Main value proposition and target customer segment, 4) Positive and negative keywords that repeat in G2 reviews over the last six months. Investigate all three simultaneously using sub-agents. Create the results as two files: an Excel comparison matrix and a Word report."

② In the Cowork sidebar, you see three sub-agents executing simultaneously. They are labeled as handling Replit, Lovable, and Bolt respectively.

③ After about five to ten minutes, two files appear in the folder. When you open the Excel file, each competitor's prices, features, and review keywords are aligned in the same rows.

E How to Do It: Practical Application

Beyond the comparison matrix, create a full-fledged report that includes strategic implications and a positioning map.

① Enter the following prompt.

"Create a competitive intelligence report for executive briefing based on the Replit, Lovable, and Bolt data analyzed earlier. Please follow this structure.

Executive Summary with three-line key findings. Competitor Positioning Map with horizontal axis showing technical difficulty (non-developer friendly to developer-focused) and vertical axis showing price range (low-cost to high-cost). Please explain each competitor's position. Price Comparison Table comparing Free, Basic, Pro, and Enterprise tiers. SWOT Matrix showing strengths, weaknesses, opportunities, and threats for each competitor. Customer Review Sentiment Analysis with frequency of positive and negative keywords and recent trends. Implications for Our Company, including identification of positioning gaps we can target and proposed market entry strategies. Create this as a Word report and attach the original data sheet in Excel separately."

② Cowork reuses sub-agents or creates new ones to complete the task. When there are areas with insufficient data, Cowork independently determines that "additional research is needed" and explores news articles or technology blogs further. This decision-making process is displayed in the sidebar.

③ When you open the completed report, the positioning map contains descriptions of each competitor's position, and the SWOT matrix contains analyses such as "Lovable has high accessibility for non-developers but is weak in scalability."

What Machines Cannot Read

The reports Cowork creates are results of analyzing text published on web pages. Information competitors intentionally hide, hidden discount conditions not revealed on pricing pages, and the context of organizational restructuring behind job postings fall outside the machine's analytical scope.

Machines cannot determine whether a competitor's sudden price cut is due to cash flow problems or is a strategic choice to monopolize the market. While review sites show rating trends, distinguishing whether those reviews are fabricated by competitors or represent genuine user voices is the work of humans.

The true value of the report begins after data collection is complete. Whether to follow the direction indicated by the numbers arrayed in the matrix, or to intuitively read the hidden sides of the market that those numbers cannot show, depends on the decision of the person holding the report.

When These Problems Arise

(1) Unable to retrieve data from a specific competitor's website. The competitor may have applied bot protection, or it could be a dynamic page requiring JavaScript rendering. Instruct Cowork with an alternative approach: "If you cannot access that page, collect the same information from cached versions, news articles, or review sites instead."

(2) Insufficient volume of review data. Startups or B2B products may have limited public reviews. Expand the search scope by instructing: "If there are fewer than five reviews on G2, search for mentions of that product on Reddit and Hacker News."

(3) Item criteria in the Excel comparison table differ by competitor. When Company A lists "100 items per month" and Company B lists "unlimited," they cannot be placed in the same column. By instructing the prompt: "For items that cannot be compared, write them exactly as they appear in the source, and add a yellow background to those cells," you make visually distinct the places where humans must later make direct judgments.

2 Latest Industry News and Trend Briefing

Why It Is Needed

When an investor or business executive opens their eyes in the morning and turns on their smartphone, they are immediately confronted with hundreds of news headlines.

New legislation from the European Union on AI regulation, earnings announcements from semiconductor firms, and geopolitical news shaking the global supply chain are all mixed together. In this massive cacophony, the task of selecting three pieces of news that truly affect my business today and summarizing their significance in three lines is a harsh ritual that must be performed every morning.

Some people use Google Alerts.

However, Google Alerts relies on keyword matching. When you set "AI regulation," it captures only articles where the text "AI regulation" appears. An article stating "The European Union announced new compliance deadlines for high-risk systems" may be missed because the keywords do not match exactly. Cowork's LLM-based filtering judges relevance at the semantic level. Even without the words "AI regulation," it detects content related to AI regulation.

What It Does

When you set up Cowork's Scheduled Task, Cowork wakes itself at the designated time, collects news, filters it, summarizes it, creates a file, and saves it to a folder. Every day, a briefing of the same quality arrives at the same time. It does not stop even when the person is busy or unwell.

The output takes two forms. One is a text briefing, where each news item is summarized in three sentences or fewer and followed by a one-sentence implication for our business. The other is an image in the form of card news. Using a Python script, key figures and one-line summaries are rendered as visual cards.

[Note] Execution Conditions for Scheduled Tasks. Cowork's scheduled tasks run locally on the user's computer, not in the cloud.

It operates only when the computer is on and the Claude Desktop app is open. If the computer is sleeping or the app is closed, it automatically runs any missed tasks when turned back on.

If a task must run at a precise time, you should adjust the computer's power-saving settings and keep the app running at all times.

How to Do It: Basic Use

Auto-generate a text briefing that summarizes five AI industry news items every morning.

Follow these steps: ① Create a folder on the desktop called "Daily_Briefing." ② Switch to the Cowork tab in Claude Desktop and grant this folder read/write permissions. ③ Enter /schedule in the chat input box. The scheduled task creation screen opens. ④ Enter the following prompt.

"Execute the following task every morning at 8 a.m. Through web search, select the five most important news items related to the AI industry that occurred in the past 24 hours. Summarize each news item in three sentences or fewer and include the source URL. Save the file to this folder with the filename format 'YYYY-MM-DD_AI_MorningBriefing.txt'."

⑤ Cowork organizes and displays the task name, execution frequency, and task details. Click the "Schedule" button to confirm.

⑥ You can view registered scheduled tasks in the "Scheduled" tab on the left sidebar.

⑦ The next morning at 8 a.m., a text file appears in the folder.

How to Do It: Applied Examples

Specify news sources concretely and add relevance tags and implications. ① Edit an existing scheduled task or create a new one. Enter /schedule.

② Enter the following prompt.

"Execute the following task every morning at 7:30 a.m. Prioritize news sources from TechCrunch, The Verge, The Electimes, and the Bloomberg Tech section. I am a business development manager at a B2B SaaS company. Filter news from this perspective. Mark each news item with a relevance rating to our business as high, medium, or low, and attach a one-sentence implication to news items marked 'high relevance.' Save as a Word file."

③ Cowork applies semantic-based filtering to screen news based on content relevance rather than keyword matching.

How to Do It: Practical Implementation

Add card news images to the text briefing and automatically send to a Slack channel.

① Go to Settings > Connectors and confirm that the Slack connector is connected. Specify the Slack workspace and the channel to send to (for example, #morning-briefing).

② Create a scheduled task with the following prompt.

"Automatically execute the following task every morning at 7 a.m. Through web search, select the five most important news items from the past 24 hours in AI, semiconductors, and cloud infrastructure. Filter out duplicate articles and prioritize fact-based articles such as earnings announcements or official product launches over rumors. I am a venture capital analyst investing in AI technology companies. Analyze the investment implications of each news item from this perspective. Save the analysis results to a Word file named 'YYYY-MM-DD_MorningBriefing.docx' in this folder. Using a Python script, create each of the top three news items from today's briefing as a card news image (PNG). Include the news title, key figures, and a one-line implication in each card.

Send the completed three card news images to the #morning-briefing channel on Slack."

③ Cowork registers a scheduled task, and on first run, uses Python's Pillow library to generate images. Card news automatically uploads to the designated channel through the Slack connector.

④ On a weekly basis, you can add a scheduled task that aggregates keywords and topics appearing repeatedly in daily briefings and generates a separate 'This Week's Trend Report'.

Ba: What machines overlook

When machines filter and summarize news, algorithmic judgment enters the process. Within articles the machine discards as unimportant may lie subtle signals that could change the world.

What people easily overlook while reading news daily is slow change. Yesterday's and today's news may seem similar, but compared to a month ago, the landscape may have shifted. An agent can capture this change numerically on a daily basis, but determining whether this change represents a truly significant transition falls to humans.

The insights the machine provides are inferences based on past data, not prophecies affirming the future. As you flip through beautifully designed card news, maintain the habit of verifying source material whenever needed, rather than blindly trusting a world edited through the machine's perspective.

Sa: If these problems arise

(1) The task did not run at the scheduled time. The computer may have been asleep or the Claude Desktop app may have been closed. When you reopen the app, pending tasks run automatically. If daily scheduled execution is needed, set automatic wake-up for that time in your operating system's power settings.

(2) News summaries are superficial. If you do not specify a persona clearly in the prompt, Cowork summarizes from a general perspective. When you narrow your area of interest, as in 'I am an analyst with focused investment in the semiconductor sector. Focus on the earnings and supply chain changes of Nvidia, TSMC, and Samsung Electronics,' the depth of analysis improves.

(3) Korean text in card news images is corrupted. This occurs when Python's image generation library cannot find a Korean font. Specify in the prompt 'Use the Korean font Malgun Gothic installed on the system' or place a font file (.ttf) in advance in the folder.

3 Professional Literature and Academic Research

Ga: Why it's needed

Suppose a researcher investigates 'the effects of vitamin D supplementation on immune function.' Entering keywords into PubMed yields thousands of papers. It takes weeks to months to complete a literature review: verifying whether each paper is a meta-analysis or clinical trial, judging whether sample sizes are adequate, filtering for research from the past five years, and comparing papers with conflicting conclusions.

PubMed alone holds over 39 million biomedical publications. At this scale, searching with a single keyword misses a significant portion of relevant papers. Results from searching 'vitamin D supplementation' and 'cholecalciferol AND immune response' may overlap by less than half. Search scope broadens only when combining synonyms, subordinate concepts, and MeSH terms (Medical Subject Headings, the medical subject classification system managed by the U.S. National Library of Medicine). Designing such search strategies itself is a specialized skill.

Na: What it does

Cowork compresses much of this process. It decomposes the broad topic you provide into multiple sub-queries and submits them simultaneously to academic databases to collect papers. It classifies collected papers by research type and extracts research design, sample size, measurement metrics, and key findings from each, creating a structured comparison table. It distinguishes areas where conclusions align from those where they conflict and identifies research gaps requiring further study.

The final output is a scholarly review document with introduction, methodology, comparative analysis, and conclusion. Every claim ends with a citation to the referenced paper.

Cowork's Opus 4.6 model has a context window of 1 million tokens (approximately 750,000 words), making it technically possible to load dozens of full papers simultaneously and cross-analyze them.

[To know] What is the Evidence Pyramid?

In medical research, not all papers carry equal weight. Systematic reviews and meta-analyses occupy the top, randomized controlled trials (RCTs) come next, and observational studies and case reports occupy lower positions. This hierarchy is called the Evidence Pyramid. When instructing Cowork to conduct a literature review, guiding it to assign weights based on this pyramid raises the scholarly rigor of the output.

Da: How to do it: basic use

Start by having Cowork read 10 papers you have collected in a folder and create a comparison summary table.

Try it yourself

① Create a folder named 'Literature_Review' on your desktop.

② Place 10 papers you downloaded earlier in this folder.

③ Grant read/write permissions for this folder in the Cowork tab.

④ Enter the following prompt.

"Read all 10 PDF papers in this folder and extract the following information from each to create an Excel comparison table. Column structure: Authors, Publication year, Study type (RCT/observational study/meta-analysis, etc.), Number of subjects, Main intervention, Measurement metrics, Key findings, Study limitations, Citation information. When extraction is complete, save it as 'Paper_Comparison_Table.xlsx'."

⑤ Cowork reads PDFs sequentially and extracts data. For large volumes of papers, it divides sub-agents for parallel processing. ⑥ An Excel file is generated in the folder. Each paper is organized into one row.

Ra: How to do it: practical examples

Instead of providing PDFs directly, you give Cowork a topic and have it find papers through web search.

① Enter the following prompt.

"Conduct a literature review on the topic of vitamin D supplementation's effects on respiratory infection prevention. Search PubMed and Google Scholar and prioritize collecting meta-analyses and randomized controlled trials published within the past five years. Secure at least 15 papers, extract research design, sample size, dosage, and key findings from each paper's abstract to create a comparison table. If there are papers with conflicting conclusions, analyze the methodological differences explaining why the results differ."

② Cowork automatically expands the search terms.

It submits variant queries simultaneously, not just 'vitamin D supplementation AND respiratory infection' but also 'cholecalciferol AND upper respiratory tract infection' and '25-hydroxyvitamin D AND pneumonia prevention'.

③ It classifies collected papers according to the evidence pyramid and assigns high weight to meta-analyses.

Ma: How to do it: real-world application

Generate a complete literature review report and identify research gaps.

① Based on the data secured in the practical examples above, add the following prompt.

"Write a formal literature review report based on the collected paper data. Follow this structure. Introduction: research background and review objective Search strategy: databases used, search terms, inclusion and exclusion criteria Evidence summary: comparison of effects between vitamin D dosage of 1000 IU daily or less versus 2000 IU or more, differences across age groups Conflict analysis: logically explain methodological differences in papers with conflicting conclusions Research gaps: identify populations (children, Asian populations, etc.) and conditions not addressed in existing research References: organize bibliographic information of all cited papers in scholarly format Generate as a Word file and attach a separate Excel comparison table."

② Cowork separates sub-agents, with one analyzing low-dosage papers and another analyzing high-dosage papers, then the main agent synthesizes the results.

③ The completed report contains critical analysis such as 'The conclusion that daily vitamin D supplementation of 1000 IU reduced respiratory infection incidence was consistently reported in 5 RCTs, but 2 showed no significant difference. This inconsistency can be explained by differences in baseline vitamin D levels among subjects.'

F. This is not a story limited to the medical field

The same framework applies equally to law, engineering, and social sciences. When comparing international legislative examples of the AI Basic Law, if Cowork systematically collects the EU AI Act, the U.S. executive order, China's Generative AI Management Methods, and Japan's AI Guidelines and creates comparison tables, researchers can focus on analysis and interpretation. The same structure works in patent disputes when conducting prior art searches and in both patent documentation searches and technical literature reviews.

G. What machines cannot identify

Literature reviews created by Cowork are drafts. Machines are accurate in analyzing the figures and methodologies written in papers, but they cannot fully account for the complex variables of the real world in which that data was produced. It is difficult for machines to perfectly filter out research designs that are subtly skewed due to pharmaceutical company sponsorship. When an overwhelming majority supports conventional wisdom, there is a risk that machines will exclude minority hypotheses that overturn that wisdom as errors.

Evaluating the methodological validity of papers, interpreting the causes of conflicting results, and reaching final conclusions are tasks for experts. Cowork is a tool that reduces the time experts need to reach judgments, not a replacement for judgment itself. When work that used to take weeks is reduced to days, the real promise of this technology is that researchers can think more deeply in that remaining time.

H. If you encounter these problems

(1) Cowork cannot read the paper PDF. If the PDF is a scanned image (saved as a picture rather than text), OCR processing is required. Instruct it as follows: "This PDF is a scanned document. Apply OCR first and then extract the text." Low-resolution scans may have lower recognition rates.

(2) Citation format differs from requirements. Citation formats vary by academic journal. Specify the desired format in your prompt, such as APA, Vancouver, or Harvard. Instruct it specifically: "Format the references in Vancouver style with numbering."

(3) Hallucination is suspected. Cowork may cite papers that do not exist or incorrectly summarize the conclusions of actual papers. You must go through a verification process of directly checking the DOI numbers of 5 to 10 key papers listed in the results on PubMed. If you instruct in your prompt, "Include a DOI number for each citation without fail. Do not use sources that cannot be verified," the risk of hallucination decreases.

4 Building a Decision-Making Bot

A. Why It Is Needed

The marketing team leader has proposed an influencer campaign.

The proposal is to pay 30 million won to a beauty influencer with 500,000 followers for a product review. The CEO asks, "What is the expected return on investment for this campaign? What were the results when we ran similar campaigns in the past?"

The marketing team leader can answer from intuition, but it is difficult to answer with data. Past campaign results are scattered across multiple spreadsheets, and the success and failure factors have never been organized systematically.

The instincts of experienced practitioners are generally correct. However, if that instinct cannot be explained in numbers, it is difficult to have persuasive power within the organization. Converting "it feels right" into "success probability of 72% based on past similar cases," that is the role of the decision-making bot.

B. What It Does

A decision-making bot is a system that learns from past data and, when you input new project conditions, calculates the expected ROI (Return on Investment) and probability of success. Cowork implements this bot in two ways.

One is pattern analysis. Cowork reads a CSV file containing the costs, influencer profiles, and result metrics of past campaigns, and establishes statistical standards for where successes and failures diverged based on which conditions.

The other is multi-perspective scenario analysis. Rather than a single AI providing answers, within Cowork, a market analysis agent, a financial modeling agent, and a Devil's Advocate agent (whose role is to deliberately offer opposing views to find holes in decisions) run simultaneously to validate the proposal from different perspectives.

C. How to Do It: Basic Use

Begin by analyzing past campaign data to judge the probability of success for a new campaign.

Follow Along

Prepare past influencer campaign data as a CSV file. The column structure is as follows: influencer name, follower count, average views, engagement rate, content type (video/image/text), product category, payment amount, campaign duration, conversions, sales increase, ROAS (Return on Ad Spend). An analysis with 20 or more records will have higher reliability.

② Create a folder called "Decision_Making_Bot" on your desktop and place the CSV file in it.

③ Grant Cowork permission to this folder and enter the following prompt.

"The CSV file in this folder is data from our company's past influencer marketing campaigns. Analyze this data and answer the following questions. 1) What was the key variable that distinguished successful campaigns (ROAS 3 times or more) from failed campaigns? 2) Was there a correlation between follower count and ROAS? 3) Were there performance differences by content type? Create the analysis results as a Word summary report and an Excel detailed analysis table."

④ Cowork reads the data and performs statistical analysis. It uses Python's pandas and scikit-learn libraries to extract correlations and patterns between variables.

⑤ The report includes analysis such as: "Engagement rate showed a stronger correlation with ROAS than follower count itself. Micro-influencers in the 100,000 to 300,000 follower range showed the highest average ROAS of 4.2 times."

D. How to Do It: Application Examples

Based on the analysis results, create a predictive model and input new campaign conditions to forecast results.

① Enter the following prompt.

"Based on the data analyzed earlier, create a predictive model that calculates expected ROAS and success probability when new campaign conditions are input. Input variables: follower count, engagement rate, content type, product category, budget. Output: expected ROAS range, success probability (probability of ROAS 3 times or higher), risk factors. Once the model is complete, test it with the following conditions: 500,000 followers, 3.2% engagement rate, video review, beauty category, 30 million won budget."

② Cowork constructs a regression analysis or decision tree model and presents test results. The format is something like: "Expected ROAS 2.8 to 4.1 times, success probability 67%, risk factor: the influencer's engagement rate has been declining over the past three months."

③ You can compare multiple scenarios by changing conditions. What if the budget is reduced to 20 million won and distributed among three micro-influencers? What if the content type changes from video to live streaming? Compare expected results side by side for each scenario.

E. How to Do It: Real-World Application

Form a team of agents and create an AI advisory committee that validates business decisions from multiple angles.

① Enter the following prompt.

"We're planning to launch a new AI Leadership Bootcamp for 7,500 dollars. To validate this decision, activate a three-agent team.

Agent 1, acting as a market analyst. Research from the web the price range, target customers, and differentiation points of competing bootcamps. Agent 2, acting as a financial modeler. Calculate the break-even point and 12-month revenue forecast assuming a price of 7,500 dollars, a target of 50 students, monthly marketing costs of 5 million won, and instructor costs of 20 million won. Agent 3, acting as a Devil's Advocate. Find at least 5 reasons why this project might fail. Dig deeply into realistic risks such as insufficient differentiation from existing free content and reduced demand for high-cost education during economic downturns. Consolidate the results from all three agents and create a one-page decision summary in Word recommending either Go, Conditional Go, or No-go."

② The three sub-agents work simultaneously. The sidebar displays the progress and intermediate results of each agent.

③ The completed summary includes recommendations like: "Initially, the 7,500 dollar price is likely to act as a barrier to entry, so we recommend a conditional go approach, starting with a 2,000 dollar short-term course and upselling the full course to graduates." Risk items identified by the Devil's Advocate agent are organized in a separate section.

F. Not Limited to Influencer Marketing

This framework is universal.

When deciding whether to enter a new market, you can apply the same structure: analyze the success and failure factors from past market entries and feed the conditions of the new market into the model. In recruitment, you can build a model to predict turnover rates or performance for a specific position based on historical hiring data.

The same applies to litigation strategy. If you accumulate data on win rates from similar cases, judicial panel tendencies, and citation rates by issue, you create a risk analysis tool that estimates the probability of winning a new case.

What Machines Cannot Guarantee

The machine's calculated success probability and projected returns rest on the premise that past data will repeat in the same pattern in the future. Events like disease outbreaks, the arrival of new competitors, and shifts in public taste are unknown variables not entered into the data.

If past data contains fewer than twenty instances, statistical reliability declines. When market conditions shift rapidly, past patterns cannot predict the future. The numbers Cowork presents are not fixed prophecies but conditional estimates. When you combine the bot's output with your own experience, grounded in that understanding, you arrive at better judgment.

At the moment you sign the final approval document before a seemingly perfect analysis dashboard, the risk and responsibility that follow from that choice remain with the person making the decision, not the machine.

If You Run Into These Problems

(1) Column names in CSV data mix Korean and English. Cowork may map columns incorrectly. If you specify the meaning of column names in the prompt,for example, 'the "followers" column is follower count, the "amount paid" column is payment amount',analysis accuracy improves.

(2) The model produces overly optimistic predictions. If historical data contains more success cases than failure cases (survival bias, when only records of those who survived remain), the model overestimates the probability of success. Instruct it: 'Considering that failure cases may be underrepresented in our data, apply a conservative adjustment to the success probability.'

(3) The devil's advocate agent is too harsh. Because its role is to intentionally find weaknesses, the result may skew negative. In the summary, treat the devil's advocate opinion as a reference item, but in your final judgment, consider the analyses of all three agents in balance.

The four tasks covered in this chapter are different facets of research. Competitive analysis maps the current landscape of the market, news briefing detects changes in that landscape daily, academic research organizes deep expert knowledge, and decision-making bots calculate future probabilities from accumulated data.

Across these four tasks, Cowork's role is the same: gather scattered information, filter out noise, and organize it into a form that allows human judgment. The role humans must take across these four tasks is also the same: doubt the results the machine has organized, examine what is absent from the text as closely as what is written in it, and make your decision on that foundation.

The startup founder in Seoul no longer needs to keep a tab open until two in the morning. When she arrives at the office in the morning, a report sits waiting in her folder. What has changed is what she does with the time she spends reading that report. Instead of copying data by hand, she thinks about what the data does not reveal. In that time of thinking lies the true value of competitive analysis.