Chapter 2: Your side of the conversation
The chapter introduction explains how working with Claude across different tasks builds confidence and identifies two foundations for using it productively: understanding the tool and understanding how prompts shape responses.
After this chapter you will be able to
- Confidence through repeated collaboration. Explain how working with Claude on different tasks can build confidence in its capabilities and help you decide when to delegate a task or use it to research and refine a plan.
- Two foundations of productive use. Describe the two things the introduction says you need to understand to work productively with Claude: the tool’s practical features and how your prompts affect the conversation and output.
- Cloud request and tool flow. Trace how a web or mobile prompt reaches Claude and where its tools run. Explain that the turn-and-response pattern and underlying models are the same across interfaces, while web and mobile work happens in Anthropic’s cloud.
- Local access and transfer limits. Explain what Claude on web and mobile cannot reach across the network boundary, and what the reader must do to provide local files or keep generated output.
- A working register with Claude. Explain what the author means by a register with Claude and why it improves with practice.
- Clear signal over polished prose. Explain why a useful prompt prioritizes relevant, specific information over elegant writing or unnecessary detail.
- What why constraints prompt pattern. Describe the section's suggested pattern for an initial prompt and what each part contributes.
- Training cutoff and recency. Explain why Claude may lack current information and what to do when a task depends on it.
- Task-specific context in prompts. Describe what task-specific details to supply to Claude and why specificity improves its output.
- Opening prompt essentials. Describe how to start a prompt by specifying the goal, relevant context, constraints, and desired output format, using details appropriate to the task.
- Clarification versus upfront detail. Explain what happened with the vague and specific cloud-migration prompts, and when the extra specificity matters most.
- Persistent conversations and session management. Explain that Claude retains past chats and their context, so you can resume an old conversation without restating your inputs, while the growing archive makes sessions harder to keep track of.
- Parallel chats and notifications. Explain how to work on multiple Claude chats at once and use notifications to know when a response is ready.
- Finding and curating chats. Describe how renaming, favoriting, and optionally deleting chats help manage a conversation archive.
- Exportable artifact outputs. Distinguish inline responses from artifacts, and describe how to request, find, inspect, download, or share artifacts.
- Restarting unproductive conversations. Explain why starting a fresh conversation with a better-framed opening prompt can beat continuing to steer a confused one.
- Human judgment of output. Explain the reader’s role in deciding whether Claude is suited to a task and whether its output is right, then iterating or changing approach accordingly.
- Example-guided output structure. Explain how showing Claude examples can help it produce a specific format, and distinguish zero-shot, one-shot, and few-shot prompting using the RFC example.
- Design decision RFC workflow. Describe how to select one design decision from an earlier conversation and ask Claude to write it as a markdown RFC using the Kubernetes Enhancement Proposal format and example KEPs.
- Synthesis of unstructured inputs. Explain how Claude can make use of unstructured inputs, such as screenshots and meeting transcripts, and how questions about their content can go beyond a summary.
- Finding and reading source material. Trace how Claude investigates a technical approach when the prompt names the blog posts but supplies no links.
- Critical technical reuse assessment. Explain what Claude was asked to judge and how the example distinguishes a promising technical pattern from permission to reuse its code.
- Mockup to design system. Describe how the author redirected Claude from building a homepage to building a design system from its mockup, and identify the resulting deliverables.
- Critical conversational steering. Explain how your next turn can accept and extend Claude’s response or stop and redirect it, and why you should judge the response against what you are trying to achieve rather than simply liking it.
- Planning and draft iteration. Describe how to iterate both before generation and after receiving output: clarify Claude’s proposed approach first, then treat the output as a draft that receives increasingly close review as it improves.
- Outdated and confident errors. Distinguish outdated information from confident incorrectness: Claude may rely on training data when current information is needed, or confidently assert a false fact such as a nonexistent command.
- Misunderstandings and critical review. Describe how Claude and the user can talk at cross purposes, and explain why the reader should notice responses that are not quite right and push back to realign.
- Evidence-based corrective feedback. Explain how to push back when Claude’s output seems wrong by using your own experience, supplying relevant evidence, and identifying factual errors precisely.
- Iteration and fresh starts. Describe when to continue correcting an unsatisfactory conversation and when to start a new one, including the options the section suggests trying.
- Coding-ready component designs. Explain why the design should be split into separate, consistently formatted component documents, with shared documentation for interfaces between systems.
- Iterative design clarification. Describe how to use the earlier system design to request deeper designs from Claude and respond when Claude identifies missing details or assumptions.
What it builds on
- Your first collaborative conversation (chapter 1) · the earlier collaborative conversation used for initial system exploration
- Choosing your model (chapter 1) · Introduces choosing a model, another option the section suggests changing.
Section by section
Introduction
The chapter introduction explains how working with Claude across different tasks builds confidence and identifies two foundations for using it productively: understanding the tool and understanding how prompts shape responses.
- Confidence through repeated collaboration: Explain how working with Claude on different tasks can build confidence in its capabilities and help you decide when to delegate a task or use it to research and refine a plan.
Why it matters: Trying varied tasks helps you make better judgments about where Claude can assist you. - Two foundations of productive use: Describe the two things the introduction says you need to understand to work productively with Claude: the tool’s practical features and how your prompts affect the conversation and output.
Why it matters: Knowing the interface alone is not enough; what you ask also shapes what Claude produces.
Key terms: delegate, research and refine a plan, practical features, prompts, conversation, output
2.1 How Claude works on web and mobile
This section explains how Claude’s web and mobile interfaces use cloud-based models and tools, and what that means for access to local data and conversation history.
- Cloud request and tool flow: Trace how a web or mobile prompt reaches Claude and where its tools run. Explain that the turn-and-response pattern and underlying models are the same across interfaces, while web and mobile work happens in Anthropic’s cloud.
Why it matters: It clarifies where your requests are processed and why a different model instance can handle the next turn. - Local access and transfer limits: Explain what Claude on web and mobile cannot reach across the network boundary, and what the reader must do to provide local files or keep generated output.
Why it matters: These boundaries determine what you need to upload and save yourself when working in web or mobile. - Cloud-stored conversation history: Describe where web and mobile chat history is stored and what that allows the reader to do across devices.
Why it matters: It explains why a conversation can be continued from another device.
Key terms: turn-and-response pattern, context of the conversation, Anthropic’s cloud, web search, sandbox environment, network boundary, upload, local network, download, chat history
2.2 Building your register
This section explains how to develop a practical way of prompting Claude: be clear and specific, give it useful context, and refine your approach through conversation.
- A working register with Claude: Explain what the author means by a register with Claude and why it improves with practice.
Why it matters: Treating prompting as a habit you develop makes it easier to improve how you work with Claude. - Clear signal over polished prose: Explain why a useful prompt prioritizes relevant, specific information over elegant writing or unnecessary detail.
Why it matters: Keeping the necessary information prominent helps Claude respond to the actual request. - What why constraints prompt pattern: Describe the section's suggested pattern for an initial prompt and what each part contributes.
Why it matters: This pattern gives Claude enough direction to produce a response you can use and refine. - Raw inputs without premature summaries: Explain why the author suggests giving Claude raw material, such as logs, instead of always summarizing it first.
Why it matters: Sharing the original material can preserve clues you might omit when describing a problem yourself.
Key terms: register, prompt engineering, clear signal, signal, noise, brevity, what, why, constraints, initial prompt
2.3 Claude’s knowledge and limits
This section explains how to use Claude’s broad technical knowledge while accounting for its limited recency, missing task context, and occasional reluctance or refusal.
- Broad technical knowledge: Explain why Claude can be useful across unfamiliar technical subjects and how discussing a familiar subject can help you judge its responses.
Why it matters: Testing Claude on a subject you know gives you a practical way to gauge its usefulness. - Training cutoff and recency: Explain why Claude may lack current information and what to do when a task depends on it.
Why it matters: A recency requirement helps Claude avoid relying on outdated information. - Task-specific context in prompts: Describe what task-specific details to supply to Claude and why specificity improves its output.
Why it matters: Relevant context reduces incorrect assumptions and unusable output. - Reluctance versus safety refusal: Distinguish Claude’s occasional reluctance about a task from its refusal of a request it considers dangerous or harmful.
Why it matters: Not every hesitant response is a final refusal, but a safety refusal is a different limit.
Key terms: training dataset, training data, web search, recency, context, prompt, supporting inputs, reluctance, Anthropic’s Usage Policy
2.4 Starting with a good first prompt
The section explains what to include in an opening prompt and why supplying relevant details upfront can avoid clarification rounds.
- Opening prompt essentials: Describe how to start a prompt by specifying the goal, relevant context, constraints, and desired output format, using details appropriate to the task.
Why it matters: These details give Claude a clearer direction from its first response. - Clarification versus upfront detail: Explain what happened with the vague and specific cloud-migration prompts, and when the extra specificity matters most.
Why it matters: Upfront detail can keep a longer task from stalling while Claude waits for you.
Key terms: goal, context, constraints, format, opening prompt, clarifications, structured version, intermediate steps
2.5 Working practically with Claude
Explains why persistent Claude chats are useful to revisit but make conversation management important as sessions accumulate.
- Persistent conversations and session management: Explain that Claude retains past chats and their context, so you can resume an old conversation without restating your inputs, while the growing archive makes sessions harder to keep track of.
Why it matters: You can return to earlier work without starting over, but you need to keep track of an expanding set of conversations.
Key terms: sessions, archive, full context, conversation management
2.5.1 Managing your conversations
This section explains how to manage Claude chats so you can work across sessions, find and revisit conversations, preserve useful exchanges, and choose when a chat should be private.
- Parallel chats and notifications: Explain how to work on multiple Claude chats at once and use notifications to know when a response is ready.
Why it matters: You need not wait in one chat while Claude works in another. - Finding and curating chats: Describe how renaming, favoriting, and optionally deleting chats help manage a conversation archive.
Why it matters: A manageable chat list makes it easier to return to work without discarding a useful record. - Summaries and conversation transcripts: Distinguish asking for a summary to resume a long chat from asking for a transcript to retain or reuse its full exchange.
Why it matters: These two requests help you pick up a conversation or reuse what happened in it. - Private chat behavior: Explain what incognito mode changes about a chat’s history and data use, including its retention caveat.
Why it matters: Knowing the limits of incognito mode helps you choose how to handle a sensitive conversation.
Key terms: chat sessions, notifications, rename, favorites, chat history, summary, transcript, incognito mode, ghost icon, safety monitoring
2.5.2 Managing Claude’s output
This section explains how to use Claude’s artifacts and how to judge, iterate on, or restart work when its output misses the mark.
- Exportable artifact outputs: Distinguish inline responses from artifacts, and describe how to request, find, inspect, download, or share artifacts.
Why it matters: Choosing an artifact makes a concrete deliverable easier to use outside the conversation. - Interactive artifact capabilities and limits: Describe what an interactive artifact does, what makes it useful for quick apps, and why it is not the place to develop a big application.
Why it matters: Interactive artifacts let you try useful app ideas quickly without treating the chat as a full development workspace. - Restarting unproductive conversations: Explain why starting a fresh conversation with a better-framed opening prompt can beat continuing to steer a confused one.
Why it matters: Restarting can save your time when continued corrections are not improving the result. - Human judgment of output: Explain the reader’s role in deciding whether Claude is suited to a task and whether its output is right, then iterating or changing approach accordingly.
Why it matters: Your value is not just obtaining output but deciding whether it serves the task and when to change direction.
Key terms: inline responses, artifacts, document icon, interactive artifacts, sandbox, Claude via the API, Claude Code, sunk cost fallacy, fresh conversation, better-framed opening prompt
2.6 Structured and unstructured data
This section explains how examples in a prompt help Claude match a desired output structure, using an RFC as an illustration.
- Example-guided output structure: Explain how showing Claude examples can help it produce a specific format, and distinguish zero-shot, one-shot, and few-shot prompting using the RFC example.
Why it matters: Examples give Claude a concrete pattern to follow when you need a particular kind of output.
Key terms: few-shot prompting, zero-shot, one-shot, RFC
2.6.1 Try it now — generating RFCs
This exercise shows how to turn a design decision from an earlier Claude conversation into an RFC by specifying a document format and providing examples.
- Design decision RFC workflow: Describe how to select one design decision from an earlier conversation and ask Claude to write it as a markdown RFC using the Kubernetes Enhancement Proposal format and example KEPs.
Why it matters: Choosing both the design question and the format gives Claude a concrete basis for a discussion-ready document. - Example guided document formats: Explain how the same approach applies beyond RFCs: give Claude your own examples, point it to public documents, or ask it to find standard examples for the kind of document you want.
Why it matters: You can reuse this way of specifying a format for many kinds of technical documents.
Key terms: RFC, Kubernetes Enhancement Proposal, KEP, markdown artifact, structured or semi-structured format, OpenAPI specs, runbooks, postmortems
2.6.2 Reading and synthesizing unstructured inputs
This section shows how Claude can turn unstructured inputs into useful outputs and answer questions grounded in their content.
- Synthesis of unstructured inputs: Explain how Claude can make use of unstructured inputs, such as screenshots and meeting transcripts, and how questions about their content can go beyond a summary.
Why it matters: You can use material you have not fully digested to ask focused questions, such as whether a technology or pattern suits your situation.
Key terms: unstructured data, queries, key themes, action items
2.6.3 Try it now — assess a technical approach
A worked example shows Claude finding blog posts and a linked GitHub repository, then critically assessing a technical approach and whether others can reuse its code.
- Finding and reading source material: Trace how Claude investigates a technical approach when the prompt names the blog posts but supplies no links.
Why it matters: The example shows how Claude can investigate unstructured material beyond what you paste into a prompt. - Critical technical reuse assessment: Explain what Claude was asked to judge and how the example distinguishes a promising technical pattern from permission to reuse its code.
Why it matters: A useful technical review considers both the merits of an approach and practical limits on reusing its code.
Key terms: unstructured inputs, GitHub repo, reference implementation, license
2.6.4 Try it now — build a design system
A Widgetario example shows how Claude can turn a homepage mockup into a design system and use that system to create more pages.
- Mockup to design system: Describe how the author redirected Claude from building a homepage to building a design system from its mockup, and identify the resulting deliverables.
Why it matters: The example shows how a clear correction can turn a proposed one-off page into reusable design assets. - Search and image extraction: Explain how Claude located the relevant mockup, extracted design information from it, and handled uncertainty about the image and requested format.
Why it matters: Finding and interpreting the actual source—and checking the format—made the generated design system more grounded. - Design system reuse: Describe how the author used the generated design files with Claude Design and what the resulting website consisted of.
Why it matters: The design system was reusable input for building consistent pages beyond the original screenshot.
Key terms: design system, DTCG format, JSON, CSS, reference HTML, deep searching, unstructured data extraction, zero-shot formatting, DTCG spec, Claude Design
2.7 Evaluating, iterating, and recovering
This section explains how to steer a conversation with Claude by critically evaluating its proposals and drafts, then iterating before investing in a thorough review.
- Critical conversational steering: Explain how your next turn can accept and extend Claude’s response or stop and redirect it, and why you should judge the response against what you are trying to achieve rather than simply liking it.
Why it matters: Your response to Claude’s first answer often determines whether the conversation moves toward your actual goal. - Planning and draft iteration: Describe how to iterate both before generation and after receiving output: clarify Claude’s proposed approach first, then treat the output as a draft that receives increasingly close review as it improves.
Why it matters: Checking the approach early and reviewing drafts progressively helps you avoid spending careful-review time on output that still needs obvious changes.
Key terms: accept and extend, block, critical approach, iterate, planning level, execution level, first draft, thorough review
2.7.1 When Claude gets it wrong
This section identifies recurring ways Claude’s responses can go wrong and explains why readers should check for subtle errors and realign the conversation when needed.
- Outdated and confident errors: Distinguish outdated information from confident incorrectness: Claude may rely on training data when current information is needed, or confidently assert a false fact such as a nonexistent command.
Why it matters: Checking both freshness and factual accuracy can catch errors that a polished answer conceals. - Blocked source access: Explain how blocked or rate-limited websites can leave Claude with incomplete source material, even when its answer draws on multiple sources.
Why it matters: An answer can be incomplete because a source was inaccessible, not because no relevant information exists. - Capability misalignment: Distinguish a real capability limit, such as lacking a required account connection, from Claude mistakenly claiming it cannot do something it can do.
Why it matters: A refusal may reflect either a genuine missing connection or an inaccurate account of what Claude can do. - Misunderstandings and critical review: Describe how Claude and the user can talk at cross purposes, and explain why the reader should notice responses that are not quite right and push back to realign.
Why it matters: Noticing a mismatch lets you correct the conversation rather than accepting an answer that misses your intent.
Key terms: Out of date information, Confident incorrectness, training data, Blocked access, rate limiting, incomplete data, Capability misalignment, Misunderstanding, angle of mistrust, push back
2.7.2 Pushing back effectively
This section explains how to challenge weak Claude outputs, provide useful corrections, and decide whether to iterate or start fresh.
- Evidence-based corrective feedback: Explain how to push back when Claude’s output seems wrong by using your own experience, supplying relevant evidence, and identifying factual errors precisely.
Why it matters: Pointing out the problem and supplying missing context can get you to a usable result—or reveal that Claude cannot help with the request. - Iteration and fresh starts: Describe when to continue correcting an unsatisfactory conversation and when to start a new one, including the options the section suggests trying.
Why it matters: You do not have to settle for the first response or keep extending a conversation that is going nowhere. - Thinking and effort tradeoffs: Explain what changing thinking mode and effort level does, and the tradeoff involved in using more effort.
Why it matters: These settings offer another way to seek a better response when extra time and tokens are worthwhile.
Key terms: push back, red herring, extra context, factual mistakes, zero-shot prompt, negative feedback, start fresh, model, thinking mode, effort level
2.8 Your turn: produce the component designs for your app
This exercise turns the Chapter 1 system design into component-level documents detailed enough to guide coding.
- Coding-ready component designs: Explain why the design should be split into separate, consistently formatted component documents, with shared documentation for interfaces between systems.
Why it matters: A coding agent needs both buildable component details and the shared interface specifications that let the pieces fit together. - Iterative design clarification: Describe how to use the earlier system design to request deeper designs from Claude and respond when Claude identifies missing details or assumptions.
Why it matters: Surfacing missing decisions before accepting the documents helps make the designs usable for implementation.
Key terms: component specifications, design documents, coding agent, interfaces between systems, initial system design, new chat session, gaps, assumptions, decision points
2.9 Summary
Where this is used later
- Why you need the desktop app (chapter 3) · Provides a baseline for understanding how the desktop app differs from web and mobile.
- Working with local files and folders (chapter 3) · Sets up the contrast with working with local files and folders.
- Your first Claude Code session (chapter 5) · The first Claude Code session introduces the separate environment the text identifies for working on application code.
- Your turn: Put your project on a maintenance schedule (chapter 4) · Working with structured and unstructured project information.
- Your turn: Set up your application project (chapter 3) · The component designs produced for the application.
Concepts explained in this chapter
- ClaudeClaude is Anthropic's AI assistant and family of large language models.
- contextContext is information supplied to a language model that frames and conditions its response, including relevant prior conversation and background details.
- promptA prompt is an input message that specifies a question, instruction, or requested task for an AI system.
- artifactA Claude artifact is a standalone generated document or other content item that can be viewed or downloaded separately from a chat.
- linguistic registerRegister is the informal set of language and interaction conventions used when addressing a particular person or audience.
- prompt patternA prompt pattern is a reusable structure for specifying a desired outcome, its purpose, and applicable constraints.
- AI modelA model is a trained computational system that processes inputs to produce outputs according to learned patterns.
- training datasetTraining data is the collection of examples used to optimize a machine-learning model's parameters.
- critical dialogueCritical dialogue is an interaction technique that actively questions and evaluates an AI system's suggestions rather than accepting them uncritically.
- dialogueA conversation is a cumulative interaction session consisting of messages exchanged between a user and an AI assistant.
- sandboxingA sandbox environment is an isolated execution environment that restricts code and tool access from the surrounding system.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
