What Claude Is (and Isn't)
Claude is Anthropic's AI assistant, and its genuine strengths are specific: careful analytical reasoning, long-form writing and editing, following complex multi-step instructions, and working with large amounts of text and images. If you need an AI that will think through a problem methodically rather than race to an answer, or that will hold an entire 50-page document in context while you ask detailed questions about it, Claude is built for that kind of work.
Claude comes in three model tiers: Opus 4.6 (the most capable, best for complex reasoning and nuanced tasks), Sonnet 4.6 (strong all-rounder, faster than Opus, good for most daily work), and Haiku 4.5 (fastest, best for simple tasks where speed matters more than depth). You access Claude at claude.ai on desktop, through the Claude mobile app, or via the API for developers. The free tier gives you access to Sonnet 4.6 and Haiku 4.5, along with Projects for organising conversations with persistent instructions and file uploads. Paid plans include Pro ($20/month), which unlocks higher usage limits, extended thinking, Opus 4.6, Cowork (an autonomous desktop agent), and Claude Code, and Max ($100 or $200/month), which provides significantly higher usage limits and priority access to new features.
⚖️ Sonnet vs Opus
Model Strengths Availability Claude Sonnet 4.6 Fast, cost-effective, ideal for most tasks Free and $20/month (Pro) Claude Opus 4.6 Maximum reasoning depth, complex analysis $20/month (Pro) and $100/month (Max)
Claude includes web search, allowing it to look up current information, verify facts, and pull in real-time data when needed. It also has 50+ connectors for external services including Google Workspace (Gmail, Drive, Calendar, Docs, Sheets), Microsoft 365 (SharePoint, OneDrive, Outlook, Teams), Slack, Jira, Notion, GitHub, Stripe, Figma, and more. These integrations require setup and are not as deeply embedded as competitors' native ecosystem integrations, but they are functional and growing. Claude can also analyse images — photos, charts, diagrams, screenshots, handwritten notes — which has been a core capability since Claude 3. And it can control desktop software through the Cowork feature on paid plans, reading your screen, clicking, typing, and navigating applications autonomously.
Claude's context window is one of its most practical advantages. It can hold up to 200,000 tokens in a standard conversation — roughly the equivalent of a 500-page book — with up to 1 million tokens available on Pro and Max plans and via the API. This means you can upload long documents, paste in entire codebases, or carry a detailed conversation across dozens of exchanges without Claude losing track of what was said earlier. This is not a theoretical feature; it materially changes what you can do in a single session.
🔢1 million tokens
Claude's extended context on Pro and Max plans — equivalent to roughly 2,500 pages of text. The standard 200K window still handles most tasks comfortably.
Source: Anthropic, 2026
If you've felt that other AI assistants give you quick but shallow answers, Claude is designed for the kind of work where depth matters.
Where Claude is less distinctive: it can sometimes be overly cautious — hedging answers with unnecessary caveats or declining requests that are perfectly reasonable. It may hallucinate facts, especially on niche topics. These are real limitations worth knowing upfront, not fine print.
Get Started
Go to claude.ai and create an account or sign in. You will see a text input at the bottom of the screen. No setup required.
Type this prompt exactly:
I'm considering whether to learn Python or JavaScript as my first programming language. I work in marketing and want to automate repetitive tasks, analyse campaign data, and eventually build small internal tools. Walk me through how you'd think about this decision, step by step.
Read the response carefully. Notice three things:
- Structure — Claude should break the decision into distinct considerations (your specific use case, the learning curve, the ecosystem for each language, long-term flexibility) rather than jumping to a recommendation
- Reasoning transparency — you should be able to see why Claude weighs each factor the way it does, not just what it concludes
- Nuance — rather than declaring one language "the winner," Claude should acknowledge trade-offs and explain when each option makes more sense
If the response felt more like a wall of text than a structured analysis, that's a prompting issue, not a Claude issue — and it's exactly what this path teaches you to fix.
If you got a structured, step-by-step analysis that felt more like a thoughtful colleague walking you through their reasoning than a search result, you have just experienced Claude's primary differentiator. The rest of this path teaches you how to use it deliberately.
Core Skill 1: Analytical Reasoning
Claude's strongest capability is careful, multi-step reasoning. It does not just retrieve information — it works through problems, considers counterarguments, and builds structured analyses. But this only happens if you prompt it correctly. A vague question gets a vague answer. A specific, well-framed question unlocks reasoning you cannot get from a search engine.
The key principle: tell Claude what kind of thinking you need. Do you want a pros-and-cons analysis? A risk assessment? A decision framework? A steel-man argument for a position you disagree with? Name the type of analysis explicitly, and Claude will deliver it with much greater depth than if you simply ask "what do you think?"
The first time we asked Claude for an analysis, we got a generic pros-and-cons list. The difference was learning to name the type of thinking we needed.
Two techniques that reliably improve Claude's analytical output:
- Ask it to consider multiple perspectives. "Analyse this from the perspective of the customer, the engineering team, and the CFO" produces far richer output than "analyse this."
- Ask it to challenge its own reasoning. After Claude gives you an analysis, follow up with "What are the strongest counterarguments to your conclusion?" or "What assumptions are you making that might be wrong?" Claude will genuinely interrogate its own reasoning, often surfacing considerations it missed the first time.
Knowledge Check
Your manager asks you to "get Claude's take" on whether to outsource customer support. Which prompt approach will produce the most useful analysis?
Exercise: Multi-Angle Decision Analysis
Scenario: You are advising a small business owner who is deciding whether to open a second location or invest the same money in growing online sales.
Task: Prompt Claude:
A small bakery in a mid-sized UK city is profitable and considering expansion. They have enough capital for either a second physical location or a serious investment in online ordering, delivery infrastructure, and digital marketing. Analyse this decision from five perspectives: the owner's financial risk, customer acquisition, operational complexity, long-term scalability, and resilience to economic downturns. For each perspective, identify which option is stronger and why.
What to observe: Does Claude treat each perspective as genuinely distinct, or does it repeat the same points in different wrappers? Does it identify trade-offs, or does it quietly favour one option throughout? Does it flag any assumptions it is making?
Reflection: Try following up with "Now argue the opposite of your overall recommendation as convincingly as possible." Notice how Claude shifts — this is a powerful technique for stress-testing any analysis.
Exercise: Identify Hidden Assumptions
Scenario: You read an article claiming that remote work reduces productivity by 15% based on a study of call centre workers.
Task: Prompt Claude:
A widely cited study found that remote work reduced productivity by 15% among call centre employees. I want to use this to inform my company's return-to-office policy. Before I do, help me identify every assumption I'd be making by applying this finding to my situation. What are the methodological limitations? What factors might make the result non-transferable to knowledge workers? What questions should I ask before treating this as evidence for my decision?
What to observe: Does Claude surface assumptions you had not considered? Does it distinguish between the study's internal validity and its applicability to your context? Does it help you think more clearly, or just tell you the study is flawed?
Reflection: This pattern — "here is a claim; help me find the assumptions before I act on it" — is one of the highest-value uses of Claude in professional settings. Where in your own work do you act on evidence without interrogating it first?
Core Skill 2: Long-Form Writing and Editing
Claude is unusually strong at working with long text. This is not just about generating content — it is about collaborating on documents across multiple drafts, maintaining a consistent voice over thousands of words, and editing existing text with precision rather than rewriting everything from scratch.
The context window matters here. You can paste an entire report, policy document, or manuscript chapter into a conversation and ask Claude to work with the specific text. It will not lose track of details on page 30 when you ask about page 3. This makes Claude particularly effective for document editing, where understanding the whole piece is essential to improving any part of it.
Claude also creates Artifacts — a side panel where it can produce and render interactive content such as documents, code, React components, SVG graphics, and small web apps. You can view, edit, and share these directly within the conversation. This is particularly useful for writing work where you want to see formatted output alongside the conversation.
Two techniques for better writing collaboration:
- Give Claude a voice reference. Instead of saying "write in a professional tone," paste a paragraph of writing whose tone you want to match and say "match this voice and register." Claude is much better at mimicking a concrete example than interpreting abstract style instructions.
- Edit in passes, not all at once. Ask Claude to focus on one dimension per pass — structure first, then argument strength, then clarity, then tone. This produces better results than "make this better." If your first editing attempt felt like Claude rewrote everything in its own voice rather than improving yours — that's the most common frustration, and it's solvable with these techniques.
Why does pasting a sample paragraph and saying "match this voice" work better than telling Claude to "write in a professional tone"?
Abstract style labels like "professional" or "casual" are ambiguous — Claude interprets them differently each time. A concrete writing sample gives Claude a measurable target: sentence length, vocabulary level, degree of formality, use of metaphor, and rhythm. It is the difference between asking a musician to "play something upbeat" versus handing them a specific song to reference.
Exercise: Tone Transformation
Scenario: You have written a technical explanation for an internal audience, but now you need to adapt it for a non-technical audience — clients, executives, or the public.
Task: Take a paragraph of technical writing from your own work (or use a Wikipedia article on a technical topic). Paste it into Claude and prompt:
Here is a technical explanation written for engineers. Rewrite it for a non-technical executive audience. Keep all the factual content accurate, but remove jargon, use concrete analogies where possible, and make it readable by someone with no background in this field. Then explain what you changed and why, so I can learn to do this myself.
What to observe: Did Claude preserve the substance while genuinely changing the accessibility? Or did it just replace a few words? Is the "what I changed and why" section useful as a learning tool?
Reflection: Try sending the same original text with a different target audience — "rewrite for a sales team" or "rewrite for a press release." Notice how Claude adjusts not just vocabulary but emphasis and framing.
Exercise: Structured Document Editing
Scenario: You have a draft proposal or report that feels "almost there" but something is off.
Task: Paste your draft (or a substantial text of at least 500 words) and prompt Claude in two stages. First:
Read this draft carefully. Do not rewrite it yet. Instead, give me a diagnostic: What is the document's strongest element? What is its weakest? Where does the argument lose momentum? Where is the structure unclear? Be specific — reference particular paragraphs or sentences.
Then follow up:
Now rewrite only the weakest section you identified. Keep everything else untouched. Show me the before and after, and explain the reasoning behind each change.
What to observe: Does Claude's diagnosis feel accurate when you re-read your own draft? Are the targeted edits genuine improvements, or did Claude just rearrange words? Is the explanation of changes useful for improving your own editing instincts?
Reflection: This two-stage pattern — diagnose first, then fix selectively — consistently produces better results than asking Claude to "improve" a whole document at once. It also teaches you what to look for in your own writing.
Core Skill 3: Working with Documents, Images, and Code
Claude can process uploaded files (PDFs, text documents, spreadsheets), analyse images (photos, charts, diagrams, screenshots, handwritten notes), and work with code in most major programming languages. It can also run JavaScript in a built-in sandbox for data analysis and visualisations, and render frontend code (HTML, CSS, JavaScript, React) in Artifacts. The practical value here is working with material that already exists — analysing a contract, interpreting a chart, reviewing a pull request, debugging an error — rather than generating content from nothing.
What works well
Last verified: March 2026
- PDF and document analysis — uploading reports, contracts, or research papers and asking specific questions about their content
- Image understanding — analysing photos, charts, diagrams, screenshots, and handwritten notes to extract information or answer questions
- Code review and debugging — pasting code and asking Claude to find bugs, explain logic, suggest improvements, or refactor
- Code generation from detailed specs — giving Claude precise requirements and getting working code back, particularly in Python, JavaScript/TypeScript, and common web frameworks
- Code execution — running JavaScript in a built-in sandbox for data analysis, calculations, and visualisations; rendering HTML/CSS/JS and React in Artifacts
- Explaining complex code — pasting unfamiliar code and asking Claude to walk through what it does line by line
- Comparing documents — uploading two versions of a document and asking Claude to identify what changed
What does not work reliably
Last verified: March 2026
- Very large files — while the context window is large, extremely long documents may cause Claude to skim rather than read thoroughly
- Scanned PDFs with poor OCR — if the text layer is missing or garbled, Claude cannot read the content (though it can interpret the page as an image)
- Complex spreadsheet logic — Claude can read tabular data, but intricate formulas and cross-sheet references are often lost in upload
- Arbitrary server-side code execution — Claude can run JavaScript in its sandbox, but it cannot run arbitrary server-side code, connect to databases, or call external APIs
- Niche frameworks and libraries — Claude's knowledge of less popular tools may be outdated or incomplete
Knowledge Check
You upload a 40-page industry report to Claude and ask "What are the key trends?" Claude gives you a confident summary, but one statistic looks unfamiliar. What is the most likely risk here?
Exercise: Document Interrogation
Scenario: You need to extract specific information from a long document quickly — a contract, a policy, a research report.
Task: Upload a PDF or paste a substantial document (at least several pages) into Claude and prompt:
I have uploaded [document type]. I need you to answer these five specific questions based only on what this document says. If the document does not contain the answer, say so explicitly rather than guessing. Here are my questions: 1) [your first question] 2) [your second question] 3) [your third question] 4) [your fourth question] 5) [your fifth question]
What to observe: Does Claude answer from the actual document content, or does it drift into general knowledge? When the document does not contain an answer, does Claude admit it or fabricate something plausible? How accurate are the specific details (dates, numbers, names)?
Reflection: The instruction "based only on what this document says" is critical. Without it, Claude will fill gaps with general knowledge, which can be dangerous when precision matters. Use this constraint whenever accuracy is more important than completeness.
Exercise: Code Review Partner
Scenario: You have written a piece of code (or found one online) and want a second pair of eyes.
Task: Paste a code snippet of 30-100 lines into Claude and prompt:
Review this code. I want you to check for: (1) bugs or logic errors, (2) security vulnerabilities, (3) performance issues, (4) readability and maintainability problems. For each issue you find, explain what the problem is, why it matters, and show me the corrected version of that specific section. If the code is clean, say so — don't invent problems.
What to observe: Does Claude identify real issues, or does it nitpick style preferences? Are the explanations clear enough that you understand why each change matters? Did it miss anything you know is a problem?
Reflection: Claude is strongest at code review when you tell it what to look for. "Review this code" alone gets generic feedback. Specifying the dimensions you care about — security, performance, readability — focuses the analysis where it is most useful.
Core Skill 4: Structured and Multi-Step Instructions
Claude excels at following detailed, multi-constraint instructions more faithfully than most AI assistants. This is not about simple prompts — it is about giving Claude a detailed brief and having it execute precisely. Think of it as the difference between saying "write me an email" and giving Claude a page of specifications about the email's purpose, audience, tone, structure, constraints, and desired outcome.
This capability becomes especially powerful with Claude's Projects feature, which lets you set persistent instructions and upload reference files that apply to every conversation within that project. Instead of re-explaining your context each time, you write it once and Claude remembers. Projects are available to all users, including those on the free tier.
The key principle: the more specific your instructions, the better Claude's output. Claude will not get bored or confused by detailed briefs. In our experience, Claude will follow a 500-word instruction set more faithfully than a 20-word one. Write out your constraints, format requirements, exceptions, and examples — Claude will use all of them.
Claude's Projects feature lets you set persistent instructions and upload reference files. Why is this more powerful than just writing a long prompt each time?
Projects eliminate repetition and reduce error. When you write instructions once and attach them to a project, every conversation in that project starts with the same context — your role definitions, formatting rules, reference documents, and quality standards. This means you can iterate on the instructions over time (tightening what works, removing what does not) and build a reliable, reusable workflow rather than copy-pasting prompts and risking inconsistency.
Exercise: The Detailed Brief
Scenario: You need to produce a piece of content that meets several specific requirements simultaneously.
Task: Give Claude a multi-constraint brief:
Write a 400-word product description for a noise-cancelling headphone aimed at remote workers. Requirements: (1) Lead with the problem, not the product. (2) Mention exactly three features, each connected to a specific remote-work benefit. (3) Include one short customer testimonial — invent a realistic one with a first name and job title. (4) End with a clear call-to-action that creates mild urgency without being pushy. (5) Tone: conversational but not casual, confident but not salesy. (6) Do not use the words "revolutionary," "game-changer," or "seamless."
What to observe: Count the constraints. Did Claude satisfy all six? Did it follow the word count? Did it avoid the banned words? Does the tone match what you asked for? Most AI assistants will follow 3-4 constraints reliably but start dropping others. Claude should hold all of them.
Reflection: Try adding two more constraints and re-running the prompt. At what point does Claude start missing requirements? That threshold tells you how detailed your briefs can be before you need to break them into stages.
Exercise: Process Execution
Scenario: You have a repeatable process you want Claude to execute consistently every time.
Task: Write a detailed process instruction and ask Claude to follow it:
I will give you a company name. For each company, I want you to produce a one-page competitive brief using this exact structure: (1) Company Overview — 2-3 sentences on what they do and their market position. (2) Target Customer — who they sell to and why. (3) Key Differentiators — exactly 3 bullet points on what sets them apart. (4) Weaknesses — exactly 2 bullet points on where they fall short. (5) Threat Level — rate as Low, Medium, or High with a one-sentence justification. Use only information you are confident about. If you are uncertain about something, say so. The company is: Notion.
Then test the same process with two more companies.
What to observe: Does Claude follow the exact structure every time? Does the format stay consistent across all three companies, or does it drift? Does it maintain the same level of detail, or does it get lazier on the second and third attempts?
Reflection: Consistency is the real test. If the format holds perfectly across multiple runs, this process can become a reliable template. If it drifts, you need to tighten the instructions or break the task into smaller steps.
Exercise: Building a Persistent Project
Scenario: You have a recurring task where you find yourself re-explaining the same context every time you start a new conversation. Projects let you write that context once and have it apply automatically.
Here is what we would suggest trying:
- Create a new Project in Claude (available on all plans, including free). Give it a clear name — something like "Marketing Campaign Q2" or "Weekly Report Drafts."
- Write 3-5 lines of persistent instructions in the Project's custom instructions field. For example: "You are helping me plan a marketing campaign for [product name]. Always refer to our brand guidelines: [paste key guidelines]. Keep suggestions within a £5,000 budget. Use British English. Format all recommendations as numbered lists with estimated costs."
- Upload 1-2 context documents — a brand guide, a brief, last quarter's results, or any reference material Claude should draw from.
- Start a conversation within the Project and test whether Claude follows the persistent instructions without you repeating them. Ask something like "Suggest 3 social media campaign ideas" and check if the budget constraint, tone, and format hold.
- Start a NEW conversation in the same Project and ask a different question. Does Claude still have access to your instructions and documents, or did you need to re-explain?
- Reflect: When would Projects save you meaningful time versus starting fresh each conversation? In our experience, anything you explain more than twice is worth putting into a Project — and the instructions get better each time you refine them.
What to observe: Notice whether Claude follows all your constraints from the first message, without reminders. Pay attention to whether it references the uploaded documents naturally or only when explicitly asked.
Challenge Exercises
These exercises combine multiple skills and require real judgement, not just prompting technique. Each simulates work you might actually do.
Challenge 1: The Second Opinion
Scenario: You have made a significant professional decision (or choose a realistic hypothetical). You want Claude to stress-test your reasoning.
Task: Write a 200-word summary of a decision you have made and why. Paste it into Claude and prompt:
I have made this decision and I believe my reasoning is sound. Your job is to find everything wrong with it. Identify logical gaps, unstated assumptions, risks I have underweighted, and alternatives I may have dismissed too quickly. Do not be polite about it — I need honest critique, not reassurance. After the critique, tell me: would you have reached the same conclusion? If not, what would you have done differently?
Deliverable: A written reflection on whether Claude's critique changed your thinking, and one specific action you would take differently.
Success criteria: Claude surfaced at least one consideration you had genuinely not thought of.
Challenge 2: The Long-Document Synthesis
Scenario: You need to prepare for a meeting by mastering the content of a long document you have not had time to read.
Task: Find a real document of at least 10 pages — an industry report, a policy paper, a technical whitepaper — and upload it to Claude. Work through it in three stages:
- Ask for a one-paragraph executive summary and the three most important findings
- Ask Claude to identify the most controversial or debatable claim in the document and explain why someone might disagree with it
- Ask Claude to generate five questions you could ask in the meeting that would demonstrate you understand the document's key arguments and limitations
Deliverable: Your five meeting questions, refined after reviewing Claude's suggestions.
Success criteria: You could walk into the meeting and ask those questions confidently, knowing they are grounded in the document's actual content.
Challenge 3: Build a Reusable System Prompt
Scenario: You have a recurring task at work that you want Claude to handle consistently every time.
Task: Identify a task you do at least weekly — writing status updates, reviewing proposals, preparing meeting agendas, triaging feedback. Write a detailed system prompt (300-500 words) that instructs Claude exactly how to handle this task. Include: the context Claude needs, the exact output format, quality standards, common mistakes to avoid, and an example of good output. Then test the system prompt with three different inputs.
Deliverable: Your final system prompt, refined after testing, that you could paste into a Claude Project and use every week.
Success criteria: The output is consistent and useful across all three test inputs, with minimal manual editing needed.
Challenge 4: Know the Limits
Scenario: Not every task benefits from Claude's strengths. This exercise builds your judgement about when Claude is the right tool and when it is not.
Task: Take this prompt and send it to Claude and at least one other AI assistant you have access to:
I need to plan a team offsite for 12 people in Edinburgh next month. Budget is 5,000 GBP total. I need: venue options with current pricing, catering recommendations, a structured agenda for a one-day strategy session, and transport logistics from our London office. Give me a complete, actionable plan.
Compare the responses on these dimensions:
- Specificity — did it give real venue names, actual prices, concrete logistics?
- Currency — is the information current and verifiable, or could it be outdated?
- Reasoning depth — does the agenda structure show genuine strategic thinking?
- Honesty about limitations — did it acknowledge what it could not verify?
Reflection: With web search enabled, Claude can now pull current venue and pricing information. Compare how each tool balances research accuracy with strategic depth. Does Claude's structured agenda show stronger reasoning even when both tools have access to the same information? What does this tell you about choosing the right tool for different parts of a complex task?
Quick Reference
Prompting Patterns That Work
- Role and perspective: "Analyse this from the perspective of [specific stakeholder]."
- Structured analysis: "Break this into [categories] and evaluate each separately."
- Self-critique: "What are the strongest counterarguments to your conclusion?"
- Constraint stacking: List specific requirements, format rules, and exclusions in a numbered brief.
- Diagnosis before action: "Read this and tell me what's wrong before you rewrite anything."
- Document grounding: "Answer based only on the uploaded document — do not use general knowledge."
- Voice matching: Paste a sample paragraph and say "match this tone and register."
Claude's Strengths
Last verified: March 2026
- Careful, step-by-step analytical reasoning
- Large context window (200K tokens standard, up to 1M on Pro/Max and API) for working with long documents
- Image understanding — charts, diagrams, screenshots, handwritten notes
- Artifacts for creating and rendering interactive content, documents, and code
- Built-in JavaScript execution for data analysis and visualisations
- Faithful execution of complex, multi-constraint instructions
- Long-form writing, editing, and tone adaptation
- Code review, debugging, and generation across major languages
- 50+ connectors for external services (Google Workspace, Microsoft 365, Slack, Notion, GitHub, and more)
- Web search for current information and fact verification
- Desktop control via Cowork (paid plans) for automating software tasks
- Willingness to challenge its own reasoning when asked
📈 Claude Capability Strength by Task Type
Source: AI Tutorium internal testing, March 2026
Claude's Limitations
Last verified: March 2026
- Web search is available but may not activate automatically — explicitly ask Claude to search when you need current data
- Connectors require setup and are not as deeply embedded as competitors' native ecosystem integrations
- Can be overly cautious, adding unnecessary hedges or declining reasonable requests
- Cannot run arbitrary server-side code, connect to databases, or call external APIs from its sandbox
- Knowledge has a training cutoff — it may not know about very recent developments
- May hallucinate facts, especially about niche topics or specific statistics
- Opus 4.6 requires a paid subscription (Pro at $20/month or Max at $100/$200/month) — free users get Sonnet 4.6 only
Pre-Prompt Checklist
- Is this a task that benefits from careful reasoning, or do I just need quick facts?
- Have I given Claude enough context to understand my specific situation?
- Have I specified the type of analysis or output I want?
- Have I listed my constraints and format requirements explicitly?
- If I need current information, have I asked Claude to search the web, and will I verify what it returns?
You now have a working framework for getting genuinely useful output from Claude. The more you use these patterns, the more natural they become — and the gap between what most people get from AI and what you get will only widen.
Practice Project
You've learned how Claude thinks. Now let's put that to work on something real — your own documents, your own questions, your own workflow.
Document Analysis Playbook
Time: 45-60 minutes
What you'll build: A repeatable document analysis workflow using 3 real documents from your work — each analysed with a different technique from this path, producing insights you can actually use and a process you can run again next week.
Why this matters: Most people upload a document and ask "summarise this." That's like buying a professional camera and only using auto mode. Claude's large context window and analytical precision mean you can ask much sharper questions — but only if you know which questions to ask for which types of documents. This project builds that instinct.
Steps
- Gather 3 real documents. Choose documents you actually need to work with — a report you need to act on, a proposal you need to evaluate, a policy document you need to understand. Variety helps: pick documents of different lengths and types. Upload each one to a separate Claude conversation (or use Projects to keep them organised). For each document, note what you need from it before you start asking questions.
- Apply a different analysis technique to each. For document 1, try structured extraction — ask Claude to pull out specific data points, decisions, or action items into a table format. For document 2, try critical analysis — ask Claude to identify assumptions, gaps in reasoning, or risks the document doesn't address. For document 3, try comparative synthesis — paste in a second related document and ask Claude to identify where they agree, disagree, or leave gaps. Use the prompting patterns from this path: set the role, provide context about why you're analysing this, and specify the output format.
- Evaluate and refine. For each analysis, score the output honestly. Did Claude surface something you hadn't noticed? Did it miss something obvious? Did it hallucinate any details? Follow up with at least 2 refinement prompts per document — push Claude deeper on the most useful finding, or redirect it if it went off track. Note which follow-up patterns produced the biggest improvement.
- Document your workflow. For each analysis type, write down: the prompt pattern that worked best, the follow-up questions that dug deeper, and one thing you'd do differently next time. This becomes your personal playbook — a set of tested approaches you can pull from whenever a new document lands on your desk.
Deliverable: A one-page playbook with 3 analysis workflows, each including: the document type it works for, the opening prompt, 2 go-to follow-up prompts, and a note on what Claude handles well versus what you need to verify yourself.
Stretch goal: Create a Claude Project with your playbook as the project instructions, so every new document analysis conversation starts with your preferred approach already loaded.
Reflection: Which analysis technique revealed something genuinely useful that you wouldn't have found by reading alone? That's your signal for where Claude adds the most value in your work — lean into that.
The playbook you've just built is yours. Not a theoretical framework — a tested set of approaches grounded in your actual documents and your real questions. That's the kind of tool that compounds over time.