Ten minutes you don't get back
Anyone who uses Claude at work recognizes this: you open a new chat, paste in the same background information you pasted in last week, and only after ten to fifteen minutes are you actually working on the task. Claude isn't the problem. The entry point is.
The framework below draws on how Anthropic itself recommends using Claude professionally – and the most useful part of it isn't a new feature, but a clear model for matching the right tool to the right task.
What's the difference between Chat, Projects, and Artifacts in Claude?
Short answer: Chat is a one-off conversation for one-off tasks. Projects are a persistent workspace for recurring work, with standing instructions and a knowledge base. Artifacts are the output format for deliverables – code, reports, and documents shown as separate, editable blocks. Research is a fourth entry point for tasks that need multi-step search across several sources.
The choice between them shapes everything that happens afterward in the session, so it should be made before you write the prompt – not partway through.
Pick your entry point before you write the prompt
Chat is the default entry: an unstructured conversation that's saved to history and can be picked up later. Memory and search across past conversations can carry key context into new sessions, but Chat doesn't give you a Project's deliberate standing instructions and curated knowledge base. Chat works well for one-off questions, quick drafts, exploratory prompting, and tasks you won't repeat.
Projects are a persistent workspace built for recurring work. A Project holds three things: standing instructions (what Claude should know and do consistently in every conversation in the room), a knowledge base (documents and reference files uploaded once), and its own conversation history, separate from your global history. Conversations within a Project share instructions and knowledge base, but don't share context with each other.
Artifacts are the right output format when the result is a deliverable rather than a conversational reply – drafts, data tables, formatted reports, and code appear as separate, editable blocks next to the chat.
Research goes further than a regular web search: it runs multi-step search across multiple sources and synthesizes them, and is available on paid plans.
The example that shows how much this matters
A project coordinator wrote weekly status reports. Every Monday, she opened a new chat and typed in the same thing: project name, team structure, stakeholder list, formatting requirements, and last week's open items. Claude delivered good reports, but the session took 40 minutes – roughly 12 of which went purely to rebuilding context from scratch.
The fix was building a Project. Background, team structure, and stakeholder list went into the knowledge base. Report format and standing instructions about escalation thresholds went into the project instructions. Now she opens the Project on Monday morning and pastes in the week's updates – the context is already there, and the session takes 25 minutes. Report quality is the same; it's the setup tax that disappeared.
Rule of thumb: Ask whether the task repeats, whether the background context is the same across sessions, and whether the output format is consistent. If you answer yes to at least two of three, build the Project – the setup time pays for itself after two to three sessions.
Four capability layers – four different problems
Beyond the entry point, there are four layers that each solve a different problem, and they can be used individually or combined.
Skills are reusable procedures. Anthropic ships built-in Skills for common professional tasks – creating, editing, and analyzing Excel spreadsheets, Word documents, PowerPoint, and PDFs. Custom Skills can be added for task-specific workflows. Skills live at the account level, not inside a single Project, and Claude invokes them automatically when they're relevant – inside or outside a Project. Where a Project's instructions and knowledge base are its own configuration, an active Skill tells Claude to follow a defined procedure for that task type, consistently, wherever it's relevant.
Two things are worth remembering about Skills. First, they reduce variation but don't eliminate it – a Skill that generates weekly reports in a consistent format will still produce different text each time. Review remains part of the workflow no matter how well the Skill is configured. Second, Skills require a trust assessment: a Skill has access to everything you grant Claude access to in the session, so review the source and permissions before enabling a third-party Skill. Anthropic-provided Skills and Skills approved by your own organization are the low-risk starting point.
Code Execution is Claude's sandbox environment for calculations. Claude writes and runs code internally and returns the result – you don't write the code yourself. This matters because Claude generates prose by producing the most likely next sequence of text, which works well for language tasks, but for calculations it can produce plausible numbers that aren't necessarily correct. Code Execution gives you a verified result because the calculation actually runs. Use it instead of asking Claude to calculate whenever numeric output will be used or reported, when data needs to be transformed or cleaned, when the output should be a chart, or when the output should be an actual downloadable file such as .xlsx, .pptx, .docx, or .pdf.
Memory retains job-relevant facts across sessions – role context, output-format preferences, the names of standing collaborators, standing frameworks that apply across projects – so you don't have to re-enter the same context every time. Memory is most useful when actively curated: review saved memories regularly, at least monthly for active users, delete or update what's no longer accurate, and keep the set focused on what actually recurs. Project-scoped Memory keeps context separated per Project, so context from Client A doesn't surface in sessions for Client B – separate workstreams should have separate Projects. Incognito mode keeps a session out of Memory and chat history (applies to standalone chats outside Projects), useful for sensitive conversations, but it doesn't override your organization's underlying retention rules.
All four layers in one workflow
A business analyst who produced a monthly regulatory report illustrates how the layers work together. The task was consistent – take the month's regulatory updates, identify what applied to the portfolio, summarize the implications, and format to a fixed template – but high-effort. For the first two months, she ran everything in Chat: uploaded documents, re-pasted portfolio context, re-typed format instructions, and manually verified every single number. She found two errors in month one and one in month two, all before the report went out.
In month three, she rebuilt the workflow using the capability layers: portfolio context and standing format instructions into the Project, past reports into the knowledge base, a Skill enabled for the report format, and numeric calculations moved to Code Execution. Time per session went from 65 to 30 minutes – and the verification step still ran. No errors were found from month three through eight.
Four questions summarize how to choose layers: Which parts of the task are the same every time? That points to standing instructions and a Skill. What reference material recurs across sessions? That points to the knowledge base. Which outputs need to be calculated correctly, not just "sound right"? That points to Code Execution. What context do you want carried across sessions without re-entering it? That points to Memory.
Model choice: speed versus capability
Claude comes in three tiers, and choosing between them is a tradeoff between speed and capability.
Haiku is the fastest and most efficient model. It handles structured tasks well – classification, extraction, formatting, simple summarization, and high-volume routine work where speed matters most and the cost of an imperfect result is low. When a task runs at volume across hundreds of items in sequence, Haiku's speed advantage compounds.
Sonnet is the balanced tier, handling the breadth of professional tasks with solid quality: drafting, synthesis, analysis, research assistance, and document review. For most knowledge work, Sonnet is the right starting point.
Opus is the higher-capability tier, for tasks that require nuanced judgment, complex multi-step reasoning, ambiguous input that needs interpretation, or work where quality outweighs speed – client-facing deliverables, complex document analysis, strategic planning, and high-stakes synthesis across multiple sources.
In practice: reserve Opus for work where the quality ceiling actually matters, use Haiku where volume and speed are primary and the task has clear structure, and let Sonnet cover the rest. On metered or quota-based plans, including API access, this tradeoff also becomes a cost tradeoff per call, since a higher-capability tier consumes more quota per call.
Context is a budget
Model choice and capability configuration determine what Claude can do in a session. Context management determines how long Claude keeps doing it well.
Every conversation has a finite working memory, and it depletes as the conversation grows. As it nears the limit, claude.ai automatically summarizes earlier messages to free up space (on paid plans with Code Execution enabled), and the full history remains available for reference. In practice: a long session where you gave detailed instructions in the first ten minutes can produce answers 90 minutes later that don't follow them – not because Claude is ignoring them, but because details can get compressed away when early context is summarized.
Signs a conversation needs an intervention: Claude stops following instructions it correctly followed earlier in the same session, replies address only the latest exchange with no reference to earlier decisions, or accuracy drops in ways consistent with lost early context.
Three responses when context degrades:
- Restart. Start a new conversation – a Project's standing instructions and knowledge base carry over automatically, the conversation thread doesn't. Right when the session has drifted past what can be salvaged, or when you're starting a genuinely new task within the same workstream.
- Summarize. Before starting a new conversation, ask Claude to produce a summary of current status – decisions made, work in progress, open questions – and paste it in as context in the new conversation. This preserves continuity without carrying a degraded context window forward.
- Preserve. For information that should be available in all future sessions, save it to Memory or update the Project's knowledge base. It's more efficient to store the right information at the right time than to re-enter it repeatedly.
Claude's usage limits operate on more than one time window – a short rolling session window, plus weekly limits on paid plans that apply across models (with a separate weekly pool for the highest model tier). The specific windows and quotas vary by plan and change over time, so check current limits in Anthropic's help center as you read this. For intensive tasks, it's more efficient to plan ahead than to work around an interrupted session: break large tasks into segments, save intermediate results to the knowledge base, and restart from a summary rather than stretching one session indefinitely.
Five things that hold up
- Pick your entry point before you write the prompt – Chat for one-off work, Projects for recurring work with stable context, Artifacts for deliverables, Research for tasks that need current information from multiple sources.
- Four capability layers solve four different problems – Projects carry context, Skills define recurring procedures, Code Execution verifies calculations, Memory preserves continuity.
- Model tiers reflect a tradeoff between speed and capability – match the tier to what the task requires.
- Context is a budget – long conversations degrade as context fills up; restart, summarize, or preserve is more efficient than working around drift after it's accumulated.
- Variation is inherent, review is structural – every Claude feature, including configured Skills, produces a different result from run to run. The framework for the right entry point and capability layer determines how work is set up; review discipline determines what actually goes out to the client.
What IT Buddy recommends
Most businesses we meet use Claude or ChatGPT in the browser, alongside the work – not because the tool is weak, but because nobody has set up Projects, Skills, or Memory correctly from the start. It's the exact same pattern that causes Copilot to underdeliver: the groundwork is missing, not the tool's capability.
The framework above is something we set up directly with our clients as part of the delivery – proper Project structure, curated Skills, and a Memory policy that doesn't leak context between clients or departments. It's part of the groundwork an AI Ready assessment surfaces: how much time your team is actually losing to setup today, and what it takes to close the gap.
Get in touch and book a free review →
Also read: Claude vs. ChatGPT for Business