AI Learn
Anthropic · AI Tool Tutorial

Claude: step-by-step practical guide

A practical tutorial for document work, writing, Projects, Artifacts and careful coding workflows.

Reviewed September 19, 2026Independent tutorialFeatures may vary
Not official support. AI Learn is an independent education site and is not affiliated with Anthropic. Check the provider’s current documentation for plan-specific limits, pricing and feature availability.
Before you start

Claude is a conversational AI assistant. Its interface and available capabilities can change by plan and release. This guide focuses on durable workflows: giving context, working with long documents, iterating on writing, using Projects and Artifacts when available, and reviewing outputs before use.

1. Give Claude a concrete job

Start with a deliverable: “Review this proposal and produce a table of assumptions, evidence, risks and unresolved questions.” Add the audience and constraints. Avoid prompts that only say “analyze this” because the model has no definition of success.

2. Use long documents with a question framework

For a report, contract, transcript or research packet, specify which parts matter and what evidence format you want. Ask for section-level references where useful. If multiple documents disagree, require Claude to show the conflict rather than silently merge them.

3. Iterate on writing with explicit editing passes

Separate drafting from editing. First ask for structure, then a draft, then a specific edit pass such as clarity, brevity, tone or fact-check flags. This is easier to control than asking for “perfect writing” in one request.

4. Use Projects for stable context

When Projects are available, group related chats and reference material around one durable objective. Add style guides, background documents or recurring constraints once, then reuse them across conversations. Keep project instructions short enough that you can audit them.

5. Use Artifacts for work you need to inspect

When Artifacts are available, ask Claude to create a document, code snippet, diagram or other work product that benefits from a separate editable/previewable surface. Treat the artifact as a draft you can inspect and revise, not an automatically approved final result.

6. Review coding output like a pull request

For code, define language, runtime, inputs, expected behavior, failure cases and tests. Ask for the smallest change that solves the problem. Before running generated code, inspect dependencies, file writes, network calls, credentials and destructive commands.

Reusable prompt pattern

Task → source/context → audience → constraints → output structure → uncertainty rule. Example: “Using only the attached brief, draft a one-page launch memo for executives. Separate confirmed facts from assumptions and list missing information at the end.”

Common mistakes

Feeding a large document set without telling Claude which source is authoritative; asking it to preserve every detail while also making text much shorter; treating a polished paragraph as factually verified; and executing generated code without review.

What to practice next

Take one real document you understand well. Ask Claude for a summary, then compare it with the source. Rewrite the prompt so every missed or distorted point is addressed by an explicit instruction. That exercise teaches you how to build a reliable review prompt.

开始之前

Claude 是对话式 AI 助手,界面和可用能力会随套餐与版本变化。本教程重点讲更稳定的工作方法:如何提供上下文、处理长文档、分阶段改写、在可用时使用 Projects 与 Artifacts,以及在真正使用结果前进行复核。

1. 给 Claude 一个具体工作

先定义交付物。例如:“审核这份方案,输出一个表格,包含假设、证据、风险和未解决问题。”再补充受众和约束。只说“分析一下”通常不够,因为模型不知道什么才算完成。

2. 处理长文档时先定义问题框架

面对报告、合同、访谈记录或研究资料包,先说明哪些部分最重要、希望怎样呈现证据。必要时要求章节级依据。如果多份文档互相冲突,要要求 Claude 明确列出冲突,而不是悄悄把不同说法合并。

3. 写作任务分阶段完成

把起草和编辑分开:先要结构,再写草稿,然后分别做清晰度、精简、语气或事实核验标记。相比“一次写到完美”,这种方式更容易控制。

4. 用 Projects 保存稳定上下文

账号支持 Projects 时,把同一长期目标相关的对话和资料放在一起。可以加入风格指南、背景资料和长期约束,然后在多个对话中复用。项目级指令不要写得过长,保持自己能够检查和理解。

5. 用 Artifacts 承载需要检查的成果

支持 Artifacts 时,可以让 Claude 生成文档、代码、图示或其他适合单独预览和编辑的成果。把 Artifact 当成可以检查和迭代的草稿,而不是自动批准的最终成品。

6. 把生成代码当成 Pull Request 来审

写代码时先指定语言、运行环境、输入、预期行为、失败情况和测试。尽量让模型做最小改动。运行前检查依赖、文件写入、联网操作、凭据和破坏性命令。

可复用提示词结构

任务 → 来源/上下文 → 受众 → 约束 → 输出结构 → 不确定性规则。例如:“只使用附件简报,为管理层起草一页发布备忘录。把已确认事实与假设分开,并在最后列出缺失信息。”

常见错误

一次塞入大量文档却不说明哪份更权威;既要求“保留所有细节”又要求“大幅缩短”;把文字写得流畅当成事实已经核验;不检查就直接运行生成代码。

下一步练习

找一份你非常熟悉的真实文档,让 Claude 做摘要,然后逐条和原文对照。把它漏掉或扭曲的内容转化成明确指令,再重新运行。这个练习能快速训练可靠的文档提示词。

開始之前

Claude 是對話式 AI 助手,介面和可用能力會隨方案與版本變化。本教學重點講更穩定的工作方法:如何提供上下文、處理長文件、分階段改寫、在可用時使用 Projects 與 Artifacts,以及在真正使用結果前進行複核。

1. 給 Claude 一個具體工作

先定義交付物。例如:「審核這份方案,輸出一個表格,包含假設、證據、風險和未解決問題。」再補充受眾和限制。只說「分析一下」通常不夠,因為模型不知道什麼才算完成。

2. 處理長文件時先定義問題框架

面對報告、合約、訪談記錄或研究資料包,先說明哪些部分最重要、希望怎樣呈現證據。必要時要求章節級依據。如果多份文件互相衝突,要要求 Claude 明確列出衝突,而不是悄悄把不同說法合併。

3. 寫作任務分階段完成

把起草和編輯分開:先要結構,再寫草稿,然後分別做清晰度、精簡、語氣或事實核驗標記。相比「一次寫到完美」,這種方式更容易控制。

4. 用 Projects 保存穩定上下文

帳號支援 Projects 時,把同一長期目標相關的對話和資料放在一起。可以加入風格指南、背景資料和長期限制,然後在多個對話中複用。專案級指令不要寫得過長,保持自己能夠檢查和理解。

5. 用 Artifacts 承載需要檢查的成果

支援 Artifacts 時,可以讓 Claude 生成文件、程式碼、圖示或其他適合單獨預覽和編輯的成果。把 Artifact 當成可以檢查和迭代的草稿,而不是自動批准的最終成品。

6. 把生成程式碼當成 Pull Request 來審

寫程式時先指定語言、執行環境、輸入、預期行為、失敗情況和測試。盡量讓模型做最小改動。執行前檢查依賴、檔案寫入、聯網操作、憑據和破壞性命令。

可複用提示詞結構

任務 → 來源/上下文 → 受眾 → 限制 → 輸出結構 → 不確定性規則。例如:「只使用附件簡報,為管理層起草一頁發布備忘錄。把已確認事實與假設分開,並在最後列出缺失資訊。」

常見錯誤

一次塞入大量文件卻不說明哪份更權威;既要求「保留所有細節」又要求「大幅縮短」;把文字寫得流暢當成事實已經核驗;不檢查就直接執行生成程式碼。

下一步練習

找一份你非常熟悉的真實文件,讓 Claude 做摘要,然後逐條和原文對照。把它漏掉或扭曲的內容轉化成明確指令,再重新執行。這個練習能快速訓練可靠的文件提示詞。

Official references

Check current product details at the source

Build the transferable skill

Prompt Engineering

After learning the interface, learn the prompt structure that works across tools.

Continue →