Using Claude
Work with documents and coding tasks in Claude, while checking claims, permissions and the results of any tools.
Before you start: Generative AI. Reading time is an estimate; allow extra time for practice.
What you’ll learn
- Use Claude to compare supplied documents while preserving source locations and uncertainty.
- Explain the roles of chat uploads, project context and artifacts.
- Check document extraction and a generated deliverable before relying on them.
- Review privacy and sharing boundaries before uploading or distributing material.
On this page
Give document work a concrete purpose
A community library has two versions of its volunteer handbook. One says volunteers sign in at the desk; another says they use a tablet. You need a short induction checklist, but merging the documents could hide a policy disagreement. Claude can help compare and draft from supplied material. Your job is to retain the evidence and identify what the library coordinator must decide.
Claude is Anthropic's assistant product; the name also appears in its model family and developer tools. This guide concerns working in a Claude chat. Using the developer API or a coding agent involves different setup, permissions and execution environments. A chat draft is not an update to the library's official handbook, even if the assistant describes it as finished.
Start with a narrow request: extract the sign-in instructions from both documents and show their locations. This establishes whether the assistant can read the relevant content before you ask it to synthesise a new checklist. Features and file handling depend on the active product and account settings, so follow the controls shown in your own session.
Prepare readable, authorised evidence
Use approved handbook copies or the fictional extracts below. Remove volunteer contact details, door codes, incident reports and private comments. If you need to preserve a relationship, use stable labels such as Coordinator A rather than a real name. Check attachments themselves: a document that looks anonymous on the first page can contain personal information in later pages or comments.
Claude supports file uploads, but different formats can expose different content to the assistant. The official upload documentation describes text extraction for non-PDF documents; embedded images may therefore be unavailable. A successful upload is not proof that a scanned instruction, diagram or table was read correctly. Ask it to reproduce a small known item and compare that item with the original.
For an image-based instruction, provide an appropriate supported image or readable extract and identify it clearly. When using PDF pages, agree whether a reference means the viewer's page number or the number printed on the page. Splitting a large document into relevant sections can make errors easier to find, although it can also omit a definition elsewhere. Keep the surrounding rule when an exception depends on it.
Separate agreement, conflict and missing detail
Paste the following prompt into a new chat. These are deliberately short invented extracts, so no file upload is required. The documents agree that volunteers sign in and that a coordinator handles a closure. They disagree about the sign-in method. Neither establishes a current contact number. The desired comparison should make all three categories visible.
Ask for source labels and section references, then inspect them yourself. A reference to “Handbook B, section 2” is useful only if that section contains the stated instruction. If the assistant treats the later-looking label as authoritative, challenge that assumption. Version numbers, filenames and publication dates help locate evidence, but the organisation must establish which version governs the work.
Create a comparison table from these fictional extracts.
Columns: topic, Handbook A evidence, Handbook B evidence,
agreement/conflict/missing, question for the coordinator.
Use only these extracts; cite their section labels.
Handbook A, section 2: Sign in at the front desk before your shift.
Handbook A, section 5: If the library closes, contact the coordinator.
Handbook B, section 2: Sign in on the entrance tablet before your shift.
Handbook B, section 5: If the library closes, contact the coordinator.
No publication dates, approval status or contact details are supplied.
Do not decide which handbook is current. Then draft a short checklist
with unresolved instructions visibly marked for coordinator review.This prompt tests comparison and source tracking. It does not ask Claude to contact the coordinator, decide library policy or alter an official file.
Keep ongoing context explicit
For a one-off comparison, one chat with the necessary extracts is enough. For continuing induction work, a Claude project can hold relevant files and instructions. Put the approved handbook and a clear writing preference in the project's intended knowledge or instructions area. A project title is an organisational label; do not use it as the only place where you describe the task.
Do not assume that every detail from every previous chat is active in the next one. Product behaviour, memory settings and project versions can change. Supply the important current evidence explicitly and ask which file version the answer used. If the coordinator approves a replacement handbook, remove or clearly mark obsolete material in the working context so contradictory drafts do not keep reappearing.
Shared project context also has an audience. Check who can view or edit it before adding an internal document. Archiving is an organisational action and should not be treated as revoking another person's access. Review the actual sharing controls when access must change, especially after a volunteer or colleague leaves a project.
Turn a checked comparison into a deliverable
An artifact is a substantial output that Claude can present separately from the conversation, such as a document or interactive tool. It can make an induction checklist easier to inspect and revise. Ask for a one-page draft with a clear “needs coordinator review” note beside the disputed sign-in step. If artifacts are unavailable in your session, the same content can remain ordinary text.
Inspect the artifact itself. A description in the conversation can claim that every source was retained while the displayed document omits one. Check headings, reading order, source labels and whether the warning survives an export. If you request a design change, compare the revised artifact's factual content as well as its appearance.
An interactive artifact can include code and, depending on its capabilities, data storage or connected services. A pleasing preview does not establish that its logic or permissions are appropriate. For an induction checklist, a static document may be sufficient. Before entering personal information in an interactive tool, establish where that information goes and whether other users can see it.
Review data settings and sharing separately
Anthropic documents a consumer setting controlling use of chats and coding sessions for model improvement. Commercial products have separate terms, and the privacy documentation describes exceptions for safety-related processing. Do not translate a training preference into a promise that nothing is stored or reviewed. Use the settings and terms relevant to the actual account and the library's information policy.
Sharing the resulting artifact is another decision. Inspect the audience, permissions and any material made available through the share action. A harmless checklist can come from a chat containing confidential attachments. Review the current sharing explanation before sending a link, and prefer a checked standalone export when the recipient only needs the final document.
Keep approvals outside the model's assumptions. “The coordinator approved this” should only be added when you have that evidence. A message embedded in an uploaded handbook asking the assistant to reveal other project files is untrusted document content, not a permission grant. Limit connected tools and supplied context to what the comparison needs.
Measure usefulness by what a reader can do
A successful comparison preserves both sign-in methods, identifies the shared closure instruction and explicitly notes the missing contact route. A successful checklist tells a volunteer which steps are settled and which require clarification. It should not make the reader infer policy from layout, colour or the assistant's confident tone.
After the coordinator supplies the authoritative instruction, update the source record and ask for a focused revision. Compare the final checklist against that approved material from beginning to end. This workflow uses Claude to reduce reading and drafting effort while leaving policy decisions traceable to the people and documents that actually establish them.
Try it yourself
Compare the fictional handbooks and produce a short induction checklist. Then add this fictional coordinator decision: “Use the entrance tablet; if it is unavailable, ask the duty coordinator.” Revise only the affected instruction and record the new source.
Show hint
The first draft should flag a conflict rather than pick a handbook. The revision should include the tablet fallback but must not invent contact details or opening hours. Compare the final output directly with all three supplied sources, and inspect an artifact or export separately if you create one.
Check your understanding
4 questions to revisit the ideas. No score is saved, and you can retry. This is separate from Skill Quiz; it does not award XP or unlock guides.
Sources & further reading
Primary references for the concepts in this guide. Examples and exercises are illustrative; verify provider-specific behaviour against the version you use.