Use Claude for Responsible, Evidence-Based Work: A Beginner Course

Use Claude for Responsible, Evidence-Based Work: A Beginner Course

Claude can help you turn an unclear question, a draft document, or a set of meeting notes into a structured starting point. It is not a substitute for professional judgement, source verification, policy review, or accountability. This course adapts the supplied How to Use Claude: A Complete Beginner’s Practical Tutorial into ByteNib’s learning-first format. It is an independent educational guide, not an Anthropic product manual or endorsement. Product features, models, availability, and limits can change, so use the official Claude guidance linked below before relying on a feature in a live workflow.

Complete source course pack: The supplied 57-page course, including its original exercises, model answers, screenshots, diagrams, troubleshooting tables, glossary, and reference list, is available as a downloadable PDF course pack. The online course below preserves every substantive learning area in a web-first format, while the PDF retains the source document in full.

Course format: approximately 20 minutes of guided reading plus 3–4 hours for the practice exercises and capstone. Level: Beginner. Primary environment: Claude in a current browser. A free account can support the core exercises, although available features and usage limits vary by plan, region, account settings, and product rollout ^[1].

What you will achieve

By completing this course, you will be able to:

• Explain the difference between the Claude application, a Claude model, a prompt, a chat, a Project, and an Artifact.
• Write a clear, reusable prompt that states the task, context, constraints, audience, and desired output.
• Improve a weak answer with precise follow-up instructions rather than restarting without direction.
• Upload a non-sensitive document and request a source-bound summary with page or section references.
• Use web search for time-sensitive questions and verify the cited source material before acting on it.
• Create a focused Project with clear instructions and an appropriate knowledge boundary.
• Turn sanitized meeting notes into a reviewable action-plan draft with owners, deadlines, assumptions, and open questions.
• Recognise hallucinations, privacy risks, ambiguous instructions, and overreliance before they become a business decision.

Before you begin

You need a current browser, a stable internet connection, and access to Claude in a supported location. Anthropic states that Claude users must be at least 18 years old ^[1]. Use only an account and environment permitted by your organisation. Prepare a short, non-sensitive practice document or sanitized meeting notes. Do not use credentials, medical data, payment information, identity documents, confidential customer records, or an original file that you cannot replace.

Before your first exercise, review Claude’s privacy controls and your organisation’s AI-use policy. Anthropic advises consumer users to be thoughtful about sharing highly sensitive information and to adjust applicable privacy settings directly in Claude ^[2].

Course map, scope, and boundaries

This course follows the source material’s complete beginner sequence: foundations; terminology; account and privacy setup; interface orientation; first conversation; prompting; follow-up refinement; file and image analysis; web research; Projects, memory, and Artifacts; a practical action-plan project; exercises; troubleshooting; responsible AI; final checks; next learning; and a glossary. The original PDF course pack preserves its detailed screenshots, visual callouts, full answer keys, and page-by-page exercises.

The course intentionally does not teach the Claude API, Claude Code, model training, fine-tuning, autonomous agents, Model Context Protocol, enterprise administration, or detailed product comparisons. Those topics need distinct technical or organisational controls and should follow only after a learner can work safely in the consumer application.

Optional installation and account orientation

The browser experience at claude.ai is sufficient for this course. Learners who prefer desktop or mobile access should install only the official Anthropic application from the official download page or their device’s recognised app store ^[1]. Confirm the publisher before installation. If sign-in or phone verification fails, do not retry rapidly: confirm the country code, wait for the service’s stated retry interval, and use official Anthropic support rather than attempting workarounds.

On first use, orient yourself to the sidebar, chat area, attachment controls, model indicator, settings, and any available capability controls. The application is the workspace you use; the model is the system generating the response. A useful analogy is an office: the application is the office, the model is the specialist working inside it, the prompt is the work brief, and the output is the delivered draft. A Project is a bounded workspace; an Artifact is a substantial standalone deliverable you may edit or reuse.

Working-memory, recency, and plan limits

A token is a small unit of text, such as a whole short word, part of a longer word, punctuation, or a number. Long chats, large files, complex prompts, and long outputs consume available context and usage capacity. When a conversation becomes repetitive or confused, start a new chat with a concise handover summary rather than continuing indefinitely.

Models also have a training knowledge boundary. Use web search when current information matters, and do not treat a model’s general answer as current merely because it sounds confident. Free access is sufficient for the core exercises, but message allowances, models, Projects, memory, and advanced capabilities vary by plan and account settings ^{[1] [5]}. Choose a stronger model or higher effort setting only when the complexity, evidence burden, and review process justify it. More effort can improve reasoning work, but it does not remove the need to verify the result.

Privacy, deletion, and safe account controls

Follow the current privacy controls shown in your Claude account, including any settings governing model improvement and privacy. Incognito or deleted chats can reduce visible history but do not turn sensitive data into safe training material. Retention and control details can change, so check current official documentation before relying on a specific retention period. Privacy settings are a supplementary control, not a substitute for data minimisation. Do not upload passwords, tokens, personal records, payment data, health information, confidential contracts, or proprietary material unless you have explicit organisational authority and an approved environment.

Knowledge check: foundations

Before moving on, answer these questions in your own words. What is the difference between AI and generative AI? Why is Claude not a database that always retrieves a stored answer? Name three useful, low-risk tasks it can support. Finally, explain why an important response still requires checking even when it includes citations.

Safe numbered implementation sequence

Step 1: Establish a safe learning workspace

Open Claude in your browser and confirm that you are using the official Anthropic service. Start with one bounded, low-risk learning task rather than a live operational decision. A useful first prompt is:

Explain the difference between generative AI, a large language model, and a chatbot in plain English. Use one work-related example for each term. End with three questions I can use to check my understanding.

Read the answer critically. The objective is not to accept a definition because it sounds fluent. The objective is to identify whether the terms, examples, and knowledge-check questions match credible material you can independently explain.

Step 2: Write a structured prompt

For business work, use a six-part prompt structure:

1. Task: State what you want created or analysed.
2. Context: Supply the relevant facts, source material, and organisational setting.
3. Constraints: State what must not be assumed, invented, or disclosed.
4. Audience: Name the reader and their level of knowledge.
5. Output: Specify structure, length, format, and evidence expectations.
6. Review request: Ask Claude to identify uncertainty, missing information, and assumptions.

Use this practice prompt:

Task: Draft a concise action-plan outline from the meeting notes below.
Context: The notes describe an internal service-improvement meeting.
Constraints: Use only the supplied notes. Do not infer owners, dates, decisions, or commitments that are not explicit. Mark missing information as “Unconfirmed”.
Audience: A department manager.
Output: A table with decision, owner, deadline, dependency, and open question columns.
Review request: List assumptions separately and ask three follow-up questions that would improve accuracy.

Clear prompting reduces ambiguity, but it does not guarantee correctness. Treat the first output as a draft that requires review.

Prompt anatomy, natural prompts, and role limits

The six-part structure above is compatible with the compact G-C-C-F-C method: Goal, Context, Constraints, Format, Check. Use either structure consistently. A natural paragraph prompt can carry the same elements without visible labels:

Create a six-row risk assessment for the supplied service-change notes. Use only those notes. Write for an operations manager, preserve quoted evidence, identify unknown owners as “Not established,” and finish with the three questions that must be answered before approval.

Avoid a long, contradictory list of prohibitions. A small number of explicit, testable boundaries is more reliable than dozens of negative instructions. A role instruction can help establish style or perspective, but it never gives Claude a professional licence, authority to make a decision, or access to facts you did not provide. Ask it to take the role of a critical reviewer or security-aware editor, not to impersonate a lawyer, doctor, accountant, or authorised approver.

For repeated analysis, request a consistent template such as: Risk; cause; potential impact; existing control; recommended action; accountable owner; target date; source status. Then ask for the strongest case for and against a proposal. This counterargument step helps reveal confirmation bias and unsupported assumptions.

Step 3: Improve the answer through directed follow-up

Avoid vague commands such as “make it better.” Instead, identify the change required and preserve the parts that already work. For example:

Keep the current table structure. Separate confirmed decisions from proposals. Remove any deadline that is not explicitly stated in the notes. Add a short executive summary that names the two unresolved risks.

If an answer includes an unsupported claim, respond with:

Identify every statement above that is not directly supported by the supplied material. Remove unsupported claims and label any remaining uncertainty.

This is a practical discipline: state the defect, state the permitted evidence, state the expected correction, and review the revised output.

Use the four-pass refinement method

1. Generate: Request a bounded first draft.
2. Audit without revision: Ask Claude to list unsupported claims, ambiguity, conflicts, missing evidence, and assumptions. Do not ask it to rewrite yet.
3. Correct: Give targeted instructions that name the defect, the permitted source, and the required output.
4. Finalise: Compare the result with the source and make the accountable human decision outside the model.

Use the troubleshooting guide below instead of vague correction requests.

Start a new chat with a handover summary

Start a new chat when the task changes materially, the conversation becomes repetitive, the evidence set changes, or you need a clean review trail. Include seven items in the handover: objective, audience, approved source material, confirmed facts, decisions already made, unresolved questions, and the requested next output. This preserves context without carrying forward hidden errors.

Step 4: Analyse a document without losing evidence boundaries

Claude supports common document and image formats, including PDF, DOCX, CSV, TXT, HTML, JSON, XLSX, JPEG, PNG, GIF, and WebP. Anthropic publishes specific upload limits and PDF-processing constraints, which you should review before uploading a large or image-heavy file ^[3].

For a permitted, non-sensitive file, first ask Claude to identify the file and visible limitations:

List the attached file name, apparent document type, available page or section count, and any readability limitations. Do not analyse the content yet.

Then request a source-bound summary:

Use only the attached document as the source of truth. Summarise its purpose, commitments, dates, responsibilities, risks, and ambiguities. For every material statement, give the relevant page or section. Do not fill gaps with general knowledge. Mark unclear information as “Unclear in source.”

Independently inspect at least three cited locations. For any financial, contractual, security, safety, legal, or employment-related material, review the source directly and seek the appropriate qualified reviewer before action.

Run deeper document-analysis labs

Table analysis. Ask Claude to preserve row labels, units, currencies, decimal places, totals, and footnotes. Then recalculate subtotals and final totals independently. A correct-looking table is not evidence that the calculation is correct.

Two-document comparison. Attach two permitted proposals and request a comparison across scope, exclusions, implementation responsibility, one-time cost, recurring cost, escalation, support, data protection, termination, and unresolved assumptions. Require the answer to say “Not stated” when either source is silent. Silence is not agreement.

Image analysis. Keep observation and interpretation distinct. Ask first, “What is visibly present?” Then ask, “What might this represent?” For example, “The image contains three boxes connected by arrows” is an observation; “The image represents an approval process” is an interpretation. Flag unreadable labels rather than guessing.

Use a traceability matrix for consequential work

For material decisions, create the following table and complete the final column yourself:

Common file problems include unreadable scans, mismatched printed and PDF page numbers, merged spreadsheet cells, unclear file names, and low-resolution charts. Improve the source file or split it into logical sections. Do not compensate for weak source material by asking Claude to guess.

Step 5: Use web search only when freshness matters

Use web search for questions whose answer could have changed, such as current standards, prices, software releases, public events, or product documentation. Ask for credible sources and open them yourself. Anthropic explains that web-search responses include citations, but it also advises users to cross-reference important information and use authoritative sources for critical decisions ^[4].

Try this prompt:

Search the web for the current official guidance on [topic]. Prefer primary sources. Give a short comparison of the sources, cite each material claim, and list what remains uncertain or needs verification.

Do not use web search as a substitute for access control, due diligence, legal advice, clinical judgement, or security review. Switch it off when current information is not needed, particularly when it would add unnecessary usage or context.

Select the right capability

Use this seven-step citation review: identify the claim; open the cited source; check that it supports the claim; distinguish primary from commentary; record the source date; compare an independent authority where consequence is high; and state what remains uncertain. Search can fail through outdated pages, access restrictions, localisation, unsupported links, weak sources, incomplete retrieval, ranking bias, or simply a poorly specified question.

When combining external evidence with business reasoning, separate: confirmed requirements, interpretation, assumptions, recommended actions, and matters requiring legal, financial, security, or other specialist review. This makes the difference between evidence and advice visible.

Step 6: Organise repeatable work with Projects and Artifacts

Use a Project when you need a focused workspace with defined instructions, related chat history, and permitted reference material. Projects can include instructions and knowledge files, but their availability and collaboration features vary by plan and workspace ^[5].

Create a practice Project named Meeting follow-up practice. Add a short instruction such as:

Help me prepare evidence-bound meeting follow-ups. Use only supplied material for decisions, owners, and deadlines. Clearly label unknown information, assumptions, and recommended next questions.

Use an Artifact only when you need a substantial standalone output that you will edit, reuse, or share. Before sharing an Artifact, confirm whether it contains confidential information, whether its visibility is correct, and whether any persistent or shared storage is involved. Anthropic notes that capability controls and sharing behaviour vary by account type ^[6].

Use memory, when available, only for low-risk preferences such as preferred tone, output length, or formatting conventions. Do not store credentials, personal data, contract terms, financial records, or security-sensitive information as a convenience. Before connecting external services, granting an Artifact access, publishing a shared item, or accepting generated code, apply least privilege: ask what access is required, grant the minimum, test with non-sensitive data, and revoke access that is no longer required.

Hands-on capstone: meeting notes to a verified action plan

The capstone turns the course into an auditable workflow. Use only fictional, sanitized, or explicitly authorised notes. Create a practice Project and place this instruction in its Project guidance:

Use only the meeting notes supplied in this Project. Keep facts, assumptions, recommendations, and unanswered questions separate. Do not invent owners, dates, budgets, approvals, or commitments. Use ISO dates when a date is explicitly stated. Use “Not established” when the source is silent.

Capstone source pack and prompt

Create a small fictional note set covering a workshop purpose, named attendees, proposed date, provisional budget, dependencies, and open questions. Include at least two deliberate gaps, such as an unnamed owner and an unconfirmed deadline. Then use this prompt:

<meeting_notes> [Paste only the permitted practice notes here.] </meeting_notes>

Using only the notes, produce: (1) an executive summary; (2) an action tracker with action, owner, due date, dependency, source status, and open question; (3) a risk register with risk, cause, impact, existing control, action, owner, and target date; and (4) a concise confirmation email. Do not invent information. Mark absent fields as “Not established” or “Unconfirmed”. List every assumption separately.

Required deliverables and model-review standard

Your action tracker should clearly identify confirmed actions, proposed actions, and unknown ownership. Your risk register should distinguish an explicit risk from an inferred risk. Your email should ask accountable people to confirm the unresolved entries rather than presenting them as facts.

Use this 12-point review before accepting the output: confirm that every decision appears in the source; every owner is named in the source; every date is explicit; every currency and total is preserved; every missing field is labelled; no risk is presented as confirmed without evidence; source references are usable; acronyms are expanded for the audience; open questions are actionable; the requested format is complete; the account owner approves the outcome; and the final distribution contains no sensitive data.

If you want to turn the plan into an Artifact, first confirm that Code execution and file creation is enabled where required, then create it only from the reviewed, sanitized result. Treat Artifact sharing as an additional privacy and access-control decision.

Capstone debrief and model answers

An answer is correct when it preserves source facts and makes uncertainty explicit. A provisional budget remains provisional. An owner is not assigned merely because they attended the meeting. An absent event date remains unconfirmed. A strong validation prompt asks the model to identify unsupported inferences, not to defend them. If the notes cannot support a statement, the correct result is not a plausible guess. It is a clear request for clarification.

Practice exercises and self-assessment

Complete each exercise with non-sensitive fictional material. Compare your result with the stated review standard, then consult the complete source course pack for the original detailed answers.

1. Explain simply. Ask for a plain-language explanation of a technical concept with a short example and three self-check questions. Review whether the audience and format were respected.
2. Refine a prompt. Convert a vague request about a delayed software delivery into G-C-C-F-C. Add an evidence boundary and a requirement to state assumptions.
3. Classify statements. Take five statements from a practice source and label each as fact, inference, assumption, recommendation, or unknown. Explain one classification you initially got wrong.
4. Audit a document summary. Upload an approved practice file. Require page references, verify three citations, and record any mismatch in the traceability matrix.
5. Research responsibly. Use web search for a current, low-risk public topic. Compare two primary sources, record dates, and state what cannot yet be concluded.
6. Stress-test a proposal. Ask for the strongest case for and against a low-risk fictional proposal. List the evidence each argument would require before a decision.

Troubleshooting reference

Responsible AI, intellectual property, and transparency

Responsible use includes accuracy, privacy, security, fairness, intellectual-property awareness, and transparency. Check for demographic stereotypes, cultural assumptions, and biased framing, especially when an output could affect people. Use a stakeholder-impact question such as: “Whose experience, language, accessibility need, or constraint is missing from this recommendation?”

Do not use Claude to reproduce protected works beyond what you are authorised to use, and do not assume generated content is free of licensing or attribution obligations. Check the source, licence, organisational policy, and review requirement before publication. For material work, keep a transparent record of how AI assisted. A concise disclosure can state: “AI assistance was used to structure an initial draft. The final content, factual verification, and decision were reviewed by the accountable author.”

Risk-based review matrix

Final operational checklist

Before prompting, define the decision, audience, data boundary, and consequence of error. While prompting, provide context, constraints, format, and an explicit check request. Before using the output, distinguish facts from assumptions, inspect sources, test calculations, seek counterarguments, and obtain accountable review. Before sharing, remove sensitive information, check permissions and intellectual-property obligations, record material AI assistance, and distribute only the human-approved version.

Glossary

AI: computer systems performing tasks associated with human intelligence. Generative AI: AI that creates new content. Large language model (LLM): a statistical neural-network model that generates language from patterns and context. Prompt: the instruction and material supplied to the model. Response: the model’s generated output. Chat: a sequence of prompts and responses. Context: information available to the model for the current response. Token: a small unit used to process text. Knowledge cutoff: the latest period represented in a model’s training data. Hallucination: a confident but unsupported or incorrect output. Source of truth: the permitted evidence base for a task. Grounding: constraining output to supplied or retrieved evidence. Citation: a reference enabling the reader to inspect a source. Assumption: an unproven condition treated as true for planning. Inference: a conclusion drawn from evidence but not explicitly stated. Project: a bounded workspace with instructions and knowledge. Artifact: a substantial standalone output that can be edited, reused, or shared. Memory: a saved preference or contextual detail, when the feature is available. Web search: retrieval of current public information with source links. Extended thinking: additional model effort for complex reasoning, when available. RAG: retrieval-augmented generation, which uses retrieved material to inform an answer. Least privilege: granting only the minimum required access. Data minimisation: using the minimum personal or sensitive data necessary. Bias: systematic distortion that can create unfair or inaccurate outcomes. Traceability: the ability to connect a conclusion to supporting evidence. Human-in-the-loop: a person retains review or decision authority. Disclosure: transparent statement of material AI assistance.

Recommended next learning route

After this course, progress through responsible AI use-case intake, evaluation gates for retrieval-augmented assistants, and organisational privacy and security controls. Later topics may include API development, Claude Code, enterprise administration, Model Context Protocol, secure integrations, agent governance, and formal AI assurance. Study them in that order only when you can clearly explain the evidence boundary, access model, risk owner, and human approval point for the intended use case.

Step 7: Complete the capstone, from notes to a verified action plan

Use sanitized meeting notes that contain no personal, confidential, or regulated information. Work through the following sequence:

1. Ask Claude to list explicit decisions, named owners, dates, dependencies, and unresolved questions.
2. Ask it to produce an action-plan table using only those explicit items.
3. Ask it to identify every inferred or ambiguous point and move these into an Unconfirmed section.
4. Compare the table against the notes line by line. Correct discrepancies yourself.
5. Share the revised draft with the accountable meeting owner for confirmation before distributing it.

Your final output is successful only when a human owner confirms the actions and the plan retains a clear distinction between source facts, assumptions, and recommendations.

Validate the outcome

Before marking this course complete, confirm that you can:

• Explain why a prompt needs a task, context, constraints, audience, output, and review request.
• Produce a follow-up instruction that corrects a specific defect without discarding useful work.
• Request a file-based summary that cites the supplied material and labels omissions instead of guessing.
• Explain when web search is appropriate and how you would verify a cited source.
• Create a Project instruction that sets a clear evidence boundary.
• Show that the action-plan capstone separates confirmed items, unconfirmed items, and recommendations.
• Identify one type of information that you should not upload to a consumer AI account.

Common failure modes

• Treating fluent writing as evidence: A polished response can still contain unsupported claims. Open primary sources and compare material statements.
• Prompting without boundaries: If you do not state the permitted source material, Claude may combine general knowledge with the supplied text.
• Uploading sensitive material by default: A convenient file upload can create avoidable privacy, contractual, and compliance risk.
• Treating silence as agreement: If the source does not name an owner or deadline, mark it as not stated rather than creating one.
• Using a single answer for a consequential decision: Material decisions require accountable human review and, where appropriate, specialist advice.
• Assuming product capabilities are permanent: Account plans, settings, regions, availability, and limits can change. Confirm current official documentation before teaching or operationalising a feature.

Professional safeguards

• Use only content you are authorised to share and apply the stricter of your organisation’s policy or the service’s current terms and controls.
• Minimise data before uploading. Remove credentials, customer identifiers, regulated data, confidential commercial terms, and unnecessary personal information.
• Keep source facts, model suggestions, and human decisions visibly separate in your final deliverable.
• Verify important outputs against original documents and authoritative sources. Record the evidence and reviewer when a result informs a business decision.
• Treat AI output as a productivity aid, not as legal, financial, medical, employment, security, or other professional advice.

Continue your learning

After completing this beginner course, continue to Build a Secure AI Use-Case Intake and Risk Assessment. That next tutorial moves from safe individual practice to an accountable organisational process for evaluating AI use cases.

Continue exploring: AI & Machine Learning analysis, the related implementation tutorial, and the structured learning path.

References

[1] Anthropic, Get started with Claude, accessed 21 August 2026.

[2] Anthropic, I would like to input sensitive data into my chats with Claude. Who can view my conversations?, accessed 21 August 2026.

[3] Anthropic, Upload files to Claude, accessed 21 August 2026.

[4] Anthropic, Enable and use web search, accessed 21 August 2026.

[5] Anthropic, What are Projects?, accessed 21 August 2026.

[6] Anthropic, What are Artifacts and how do I use them?, accessed 21 August 2026.