Claude for Small Business
Everything you need to use Claude well in a small business — from the first prompt to full automation. Fifteen sections, read them as you need them.
Welcome — you're set up for real work now.
Thank you for getting Claude for Small Business. What you purchased is the setup knowledge most small business owners never figure out on their own — the difference between a Claude that sounds like everyone else's and one that actually knows your business, matches your voice, and saves you hours every week.
This guide covers everything: how to write prompts that work, how your Project and system prompt are structured, how connectors pull live data, how to automate recurring work, and how to get the most out of every credit you're paying Anthropic for. You don't need to read it all at once. Use it as a reference — come back to any section the moment you need it.
One thing to know up front: Claude moves fast — new models and features ship constantly, sometimes within days of each other. This playbook is built around the parts that don't change: how to prompt, how to structure your setup, how to think about which model and which tool to reach for. Where a specific version number or price appears, treat it as a September 2026 snapshot.
What's in this guide
Everything you need to use Claude well, from the first prompt to full automation.
Start with Prompting That Works, then Projects. Those two sections are where 90% of the value lives.
Everything else you can read when you need it.
What's in your setup
| ✍ | 🗂 | 📋 |
|---|---|---|
| Prompting That Works | Projects | System Prompts |
| The principles behind getting | The container that knows your | What makes your Claude different |
| useful output — not tips, real | business. Context, files, memory | from everyone else's. How to write |
| mechanics. | — all in one place. | and improve it. |
| 🔌 | 📄 | 🧠 |
| Connectors | Artifacts | Memory |
| Gmail, Drive, Calendar, | Documents, PDFs, spreadsheets | How Claude remembers across |
| QuickBooks — Claude reads your | — outputs that persist, can be | conversations — and what it |
| real tools. | edited in place, and reused. | actually controls. |
⚙
| 🤖 | ⚡ | Models |
|---|---|---|
| Cowork | Skills | Opus 5, Sonnet 5, Haiku 4.5, plus |
| Claude working autonomously on | Repeatable workflows you encode | Fable 5.1 above them — which to |
| your files — desktop, web, or | once and reuse forever — in the | use when, and when to let Claude |
| phone. No babysitting required. | browser and in Cowork. | choose. |
💡 Credit Optimization How the token limit system works and exactly how to stop hitting it.
Prompting That Works
Most people blame Claude when output comes back wrong. Usually it's the prompt. These are the real mechanics — not a list of tips, but the principles that explain why some prompts work and others don't.
Principle 1 — Claude takes you literally
Claude doesn't assume what you meant. It processes what you wrote. If you're vague, it fills the gaps with the most statistically average interpretation — which is why you get generic output when you ask generic questions.
"Write an email to my customer about the project." Claude has no customer name, no project context, no tone, no purpose. It writes a placeholder.
"Write a follow-up email to Meridian Co.. We completed phase 1 last Friday. Tone: professional but warm. Ask for sign-off on phase 2. Three short paragraphs max."
The difference isn't effort — it's specificity. The more variables you define, the less Claude has to guess.
Principle 2 — Give Claude a role, not just a task
Claude performs better when it knows who it is in a given context, not just what to do. A role activates a different register of knowledge and tone.
"Review this proposal and tell me what's wrong with it."
✓ ROLE + TASK "Act as a skeptical prospect receiving this proposal for the first time. What would make you hesitate, ask for more, or say no? Be honest."
Useful roles: skeptical customer, senior editor, financial analyst, experienced operations manager, devil's advocate. Assign the role before the task.
Principle 3 — Instruction order matters
Claude reads your prompt start to finish, but it weights the beginning and end more than the middle. Structure your prompts so the most important constraints come first or last — not buried in the middle of a paragraph.
Proven structure: Context → Role → Task → Constraints → Format. Put what you absolutely cannot compromise at the top. Put format instructions at the bottom.
Principle 4 — Tell Claude what not to do
Positive instructions define what you want. Negative instructions eliminate the most common failure modes. The best prompts use both.
Draft a proposal for [customer] for [service].
Do: match their formal tone, use our standard scope format, reference their industry specifically.
Principle 5 — Use XML tags to separate inputs
When you're giving Claude multiple pieces of information — context, instructions, content to process — XML tags help it distinguish what's what. This matters especially when pasting in emails, documents, or data.
I run a 6-person marketing consultancy. We work on retainer and project engagements.
[paste the customer email here]
Summarize what the customer is asking for and draft a reply that acknowledges their concern without committing to a change in scope.
XML tags aren't mandatory for simple prompts. But for anything complex — multi-part tasks, long documents, pasted data — they reliably improve output quality.
Principle 6 — Direct the thinking, don't just request it
The current models already reason before they answer on hard problems — that capability is built in now, not something you have to summon. What your prompt does is aim that reasoning. Telling Claude what to think through, and in what order, produces noticeably better results than leaving it open.
"Before you recommend anything, weigh the cost against the risk of doing nothing."
"Think through the customer's likely objections first, then write the pitch that answers them."
"What are the most likely things I'm getting wrong here? Start there."
The shift from "think about this" to "think about these specific things, in this order" is where the quality comes from. You're not asking Claude to reason — it will anyway — you're telling it what to reason about.
Principle 7 — Iterate, don't restart
One prompt rarely gets you to the final output. Treat Claude like a draft-and-revise process, not a one- shot answer machine. The fastest path to a great result is a decent first draft followed by targeted corrections — not rewriting the whole prompt from scratch.
"That's good — now make it shorter and remove the second paragraph."
→
→"The tone is right but it's missing urgency. Add that to the closing."
→"Good structure. Replace all the generic phrasing with something more specific to their industry."
→"That's exactly what I needed. Save the format as a template for future proposals."
Principle 8 — When to paste vs. when to use a connector
Pasting content into a prompt works fine for short pieces — an email, a paragraph, a set of notes. For recurring data — your customer list, your service descriptions, your pricing — load it into the project once and never paste it again. For live data — last week's emails, today's calendar — use a connector. Copy- pasting live data every time is the tax you pay for skipping the setup.
Principle 9 — Lead the output to lock the format
You can start Claude's answer for it. By writing the first line or two of the output you want, you set the format and skip the preamble entirely. This is one of the most useful moves almost nobody outside of power users knows.
Write the proposal. Start your response exactly like this and continue:
## Project Overview [one paragraph]
## Scope of Work [bulleted deliverables]
Claude picks up the pattern and fills it in — no "Sure, here's a proposal!" intro, no format drift, no second prompt to clean it up. Use this whenever the shape of the output matters.
Principle 10 — Make Claude critique its own draft
One-shot output is rarely the best Claude can do. The highest-leverage move in this entire guide: ask for a draft, then in the same breath ask Claude to find what's weakest about it and fix it. You get the benefit of a revision pass without doing the reviewing yourself.
Draft this proposal. Then, before you finish, identify the three weakest things about your own draft — the parts a skeptical customer would push back on — and rewrite those parts to be stronger. Show me only the final version.
This consistently beats a single pass, and it costs you one prompt instead of three. It works for proposals, emails, strategy, positioning — anything where quality matters more than speed.
THE ONE-LINE TEST Before sending any prompt, ask: If a capable new hire with zero context read this instruction, would they know exactly what to produce? If not, add the missing context.
Next: Projects. Prompting principles tell you how to communicate with Claude. Your Project is what gives Claude the permanent context to make those prompts land correctly — your business identity, your customers, your voice, your files, all loaded before you type a single word. Go to Projects →
Projects
A Project is Claude's memory container for your business. Every time you open it, Claude already knows who you are, how you communicate, who your customers are, and what you need. You never re-explain anything.
How context works — the hierarchy
Understanding this unlocks why Claude behaves differently in different places. Context loads in three layers, each overriding the previous:
| LAYER | WHAT IT IS | SCOPE | |
|---|---|---|---|
| Project instructions | Your system prompt — always active inside this | All conversations in this | |
| Persistent | project, every conversation | project | |
| Project files | On-demand | Documents, cheat sheets, briefs you've uploaded — | All conversations in this |
| Claude reads them as needed | project | ||
| Conversation context | What you've said in this specific conversation. Gone | Current conversation | |
| Session only | when the conversation ends. | only |
Always work inside your project. Regular Claude conversations have no system prompt, no files, no ℹ persistent context. They forget everything when you close the tab. Your project is what makes Claude feel configured for you.
How Projects handle your files — the part worth understanding
Project files don't all get loaded into every conversation. Claude uses a retrieval system: when you ask something, it pulls in only the relevant parts of the relevant files, rather than reading every document end to end. This is why you can load a project with dozens of briefs, proposals, and reference docs without slowing Claude down or blowing through your usage.
Two practical consequences:
More files is → fine. Specificity → helps retrieval.old proposals." Name the file or the customer when you can.
well-stocked project is an asset, not a burden — Claude only reaches for what each question needs. "Use the structure from the Meridian proposal" retrieves more precisely than "use one of my →About the context window: Each conversation has a working memory limit — 200K tokens (roughly a 500-page book) is the standard on paid plans, and the current models — Sonnet 5, Opus 5, Fable 5.1 — are built to handle up to 1M tokens. You'll rarely hit the limit in normal use, but very long conversations or huge pasted documents can fill it. When that happens, Claude now compacts the conversation — it summarizes the earliest messages to make room rather than cutting you off — but detail from early on gets fuzzier. The better fix is still simple: start a fresh conversation in the same project. Your system prompt and files reload instantly — only the back-and-forth resets.
What lives inside your Project
📋 System Prompt Your business identity, voice, customers, services — loaded automatically in every conversation. This is what makes your Claude different from anyone else's. 🔌 Connectors Gmail, Drive, Calendar, QuickBooks — live connections so Claude reads real data rather than what you paste in.
📎 Project Files Uploaded documents Claude keeps available — your cheat sheet, past proposals, customer briefs, service descriptions. Add anything you'd want a colleague to already know. 💬 Conversation History All chats inside the project are searchable. You can find past work, decisions, and outputs at any time.
Building your project over time
→New customer:Upload their brief or contract — Claude will know them by name going forward
New service:Add a description so Claude uses your exact language, not generic alternatives
→
→Useful template:Save any artifact Claude creates to the project — it becomes a permanent reference
→Brand voice:Paste in a writing sample or voice guide — Claude matches the register
Meeting notes:Drop them in the project to have Claude reference them in future conversations
→
One project per business, or one per customer? If you serve multiple customers or clients, consider separate projects per customer — each loaded with their voice, briefs, and history. Your own business project handles proposals, quotes, operations, and internal communications. The context stays clean that way.
System Prompts
The system prompt is the instruction that runs before every conversation. It's what makes Claude behave like it knows your business rather than like a generic assistant talking to a stranger. Getting this right is the most leveraged thing you can do.
Project-level vs. conversation-level
There are two places to set instructions — and they behave differently:
| TYPE | WHERE | WHEN IT APPLIES | BEST USED FOR |
|---|---|---|---|
| Project | Project → Edit | Every conversation in the | Identity, voice, customers, |
| instructions | Instructions | project, automatically | services, standing rules — anything always true |
| In-conversation | First message or early | That conversation only | Task-specific context, temporary |
| instructions | in a conversation | constraints, one-off roles |
The project instruction is always active. In-conversation instructions layer on top of it for that session only.
What a strong system prompt includes
Business identity:Name, industry, location, what you actually do
→
→Owner voice:Pulled from real writing — if you can paste in a sample email or proposal introduction, do it
→Key customers:emails, is sensitive about timelines") →Service vocabulary:How you describe your own work — use your exact language, not generic alternatives →Output preferences:Length, formality, format — how you want things delivered by default Hard rules:Words to never use, topics to avoid, tones that are off-brand →
Names, account type (retainer, project, one-off), any relationship notes ("Meridian prefers short Top recurring tasks:Stated in your own words so Claude knows what "normal" work looks like
→
What makes a system prompt fail
The generic test: If you could paste your system prompt into any other business in the same industry and it would still work, it hasn't been written specifically enough. "I run a small business" is not a system prompt. "I run a 6-person content studio in Montreal serving mid-market SaaS companies. I write in a direct, lowercase tone and never use the word 'leverage'." — that's the beginning of one.
COMMON FAILURE MODES
Hard rulesJust like in any prompt, Claude weights the start and end of your system prompt most heavily. A
→
| buried in | non-negotiable rule ("never quote a price without my approval") parked in the middle of a long | |
|---|---|---|
| the | paragraph gets followed less reliably. Put the rules that must never break at the very top or the very | |
| middle: | bottom. | |
| →Contradicting | "Be concise" in one line and "be thorough" in another confuses the model — it picks | |
| instructions: | one arbitrarily | |
| →Over-prescribing | Telling Claude to always use bullet points regardless of task produces bad output for | |
| format: | tasks that need prose | |
| →Placeholder brackets | Any [bracket] left unfilled is read literally — Claude will start calling your customers " | |
| left in: →TokenSystem prompts load on every message. Lengthy filler (rambling introductions, vague philosophy) costs bloat:usage every single time and crowds out the actual conversation. Aim for dense and specific — | Anthropic's own guidance is to keep project instructions tight. | [customer name]" |
| Stale → information:actually write — review it every 2–3 months | Customer names that are gone, services you no longer offer, tones that don't reflect how you |
How to edit your system prompt
Open your Project on claude.ai
1
Click the project name in the left sidebar to enter it.
Click "Edit Instructions" or the pencil icon
2
This opens the project instructions panel where your system prompt lives.
3 Make your changes and save Changes apply immediately to all future conversations in the project. Existing conversations are not affected. Test with a representative prompt
4
Open a new conversation in the project and run a typical task. If the output feels off — too generic, wrong tone, wrong format — the fix is almost always in the system prompt, not the task prompt.
The fastest way to improve your system prompt: When Claude gives you output that's wrong in a consistent, repeatable way, add a specific rule addressing it. "Do not open emails with 'I hope this message finds you well'" is more effective than "write naturally."
Artifacts
Artifacts are the standalone outputs Claude creates — documents, spreadsheets, PDFs, interactive pages — that appear in a separate panel beside the chat. They persist, can be downloaded, can be edited in place, and can be reused as templates.
Types of Artifacts
| TYPE | WHAT IT IS | EXAMPLE USE | |
|---|---|---|---|
| Document | Formatted text — proposals, emails, | "Write a scope of work for this project as | |
| Markdown | reports, scope of work | a clean document" | |
| Code | Code | Scripts, formulas, automations, templates | "Write a formula for my invoice tracker" |
| Spreadsheet XLSX | A working Excel-compatible file | "Build me a customer tracker with these columns" | |
| Interactive | A live page or calculator you can click | "Build an interactive pricing or quote | |
| HTML | through and interact with | calculator" | |
| A formatted, print-ready document | "Generate a clean PDF version of this report" |
Persistent artifacts — what's new
Artifacts can now store and retrieve data across sessions. This means an artifact isn't just a one-time output — it can be a living document that remembers state between uses.
Customer trackersthat update each time you add information — no re-uploading
→
Running logs(decisions, meeting notes, project status) that accumulate over time
→
→Dashboardsthat hold your data and reflect updates each time you open them
→Prompt librariesyou can add to without rebuilding the whole thing
To use persistent artifacts, ask Claude explicitly: "Create an artifact that saves this and retrieves it next time I open it." This unlocks the storage layer. For most everyday outputs (proposals, emails, reports), a standard artifact is fine.
For visual work, there's now Claude Design. Slides, one-pagers, landing-page mockups, prototypes — anything where the picture is the deliverable — lives in Claude Design, a separate canvas included with paid plans. Artifacts remain the right tool for documents, spreadsheets, and calculators; Design is where you go when you'd otherwise open Canva or Figma.
How to work with Artifacts
1 Ask for an artifact explicitly when it matters Claude will create one naturally for documents, but saying "create this as an artifact" makes the output cleaner Use the icons in the artifact panel
2
Copy sends the content to your clipboard. Download saves it as a file. The format depends on what type of artifact it is.
3 Revise it in-conversation — or edit it in place "Make this shorter," "Change the tone to formal," "Add a pricing section" — Claude updates the artifact directly without rewriting from scratch. New since June 2026: you can also highlight the exact passage you want changed inside the draft, type the change, and Claude edits right where you marked it — no re-describing which paragraph you meant.
Save good templates to your project
4
When Claude produces a format you want to reuse — a proposal or quote structure, a scope template, a brief format — upload it to your project files. It becomes the default reference for that type of document going forward.
Memory
Claude's memory is not one thing. There are three distinct mechanisms, each working differently. Knowing which one you're relying on determines whether something is reliably available or just temporarily there.
The three memory mechanisms
📋 System Prompt The most reliable memory. Written once in your project instructions. Always loaded. Always available. This is where permanent business context lives — identity, tone, customers, services.
📎 Project Files Documents you upload to the project. Claude reads them as needed within conversations. Not auto- loaded word-for-word, but available for Claude to reference. Best for longer reference material — briefs, templates, service guides.
🧠 Auto-Memory (Settings → Memory) When enabled, Claude keeps a list of individual, categorized entries — preferences, names, recurring patterns — that it reads and updates as you work. Since August 2026 every entry is listed under Topics in Settings → Memory, where you can edit or delete any of them, and memory now carries across chat and Cowork.
Auto-memory is Claude's notes, not your transcript. Claude decides what to save — you can edit the entries after the fact, but you don't control what gets written in the first place, and the entries are short. It's much better than the old daily summaries it replaced in July 2026, but still: do not rely on auto-memory for anything important. The system prompt and project files are the memory you actually control.
The right memory for the right job
WHAT YOU WANT REMEMBERED Your business identity, voice, customers A specific customer's brief or contract A proposal template you want reused A preference discovered in conversation ("always use Oxford commas") Meeting notes from today Something that came up once
WHERE TO PUT IT System prompt — always Project file — upload it Project file — save the artifact Add it to the system prompt manually Paste into a conversation or upload to the project Don't try to "memorize" it — use context when relevant
How to manage auto-memory
→Go to Settings → Memory → Topics to see every entry Claude has saved about you
→Edit or delete anything outdated or wrong — stale memories cause incorrect behavior
→Health, beliefs, and similar personal topics are excluded unless you switch on "Include sensitive topics in
memory"
→Memory is on by default on Pro and Max, andoff by default on Team plans— a Team admin has to enable it
If auto-memory is creating noise, turn it off and manage context manually through your system prompt instead
→
Auto-memory works across all conversations and across Cowork — both inside and outside your project. Use an
→
incognito chat for anything you don't want remembered
Connectors
Connectors are live bridges between Claude and your existing tools — Gmail, Google Drive, Google Calendar, QuickBooks, and more. Once connected, Claude reads your real data without you copy-pasting anything.
What's actually happening when you use a connector
When you ask Claude a question that touches a connected tool — "what did my customer email me last week?" — Claude makes a live read request to that tool, pulls the relevant data into your context window for that conversation, and uses it to answer. It doesn't store your data. It reads on demand, in-context, and the data is gone when the conversation ends.
Claude asks before it acts. Reading your data is automatic, but anything that sends, posts, changes, or ℹ deletes — sending an email, creating a calendar event, editing a record — surfaces for your confirmation first. Claude drafts; you approve. You stay in control of every outbound action.
What each connector enables
| TOOL | WHAT CLAUDE CAN DO | EXAMPLE PROMPT |
|---|---|---|
| Gmail | Read threads, find emails by sender or topic, | "Summarize my last email thread |
| summarize conversations, draft replies | with Meridian Co. and draft a follow-up asking for a decision by | |
| Google Drive | Read documents and files, reference past proposals, | "Find my last proposal for |
| quotes, and contracts, pull content by name or topic | Northgate Supply and use its structure for this new one." | |
| Read your schedule, draft prep notes for upcoming | "I have a call with a new prospect | |
| Calendar | meetings, write follow-ups after a call | tomorrow. Based on their company name, what should I prepare?" |
| TOOL | WHAT CLAUDE CAN DO | EXAMPLE PROMPT |
| QuickBooks / | Answer financial questions with your actual data — AR | "What did I invoice last month vs. |
| Xero | aging, expenses, revenue by customer or period | last month a year ago, and what's still unpaid?" |
| Notion | Read pages and databases, reference project notes, | "Check my customer database in |
| pull structured data from any Notion table | Notion and tell me which accounts are up for renewal this quarter." | |
| Microsoft | Outlook mail and calendar, OneDrive and SharePoint | "Find the SharePoint contract for |
| 365 | files, Teams (read-only). Since July 2026 it can also | Northgate and draft a reply to their |
| draft and send mail, manage events, and create files — | last Outlook email confirming the | |
| once an admin turns write access on | renewal date." |
How to add a connector
1 Go to Settings → Connectors on claude.ai Click your profile icon (top right), then Settings, then the Connectors tab. 2 Find the tool and click Connect Scroll through the list. Native connectors (Gmail, Drive, Calendar, Microsoft 365, QuickBooks, Notion, Slack, 3 Authorize via OAuth Follow the permission screen. You control what Claude can access. You can revoke access at any time from the Test it inside your project
4
Open your project and ask a question that uses the new connector. If it works, Claude will pull live data. If it doesn't, check that the connector is authorized and the tool has data in it.
Connectors are account-level, not project-level. Once you connect Gmail, it's available across all your projects. You choose whether to use it in a given conversation by simply asking a question that requires it.
Search & Deep Research
Claude can search the web in real time and run deep multi-source research on any topic — competitors, regulations, markets, industry news — and return structured output you can actually use.
Web Search
Enabled by default on paid plans. When you ask about anything current — news, recent data, a company, a regulation — Claude searches automatically and cites sources inline. You don't need to do anything to activate it.
Toggle web search on/off using the globe icon at the bottom of the chat input. Turn it off when you want ℹ Claude to work purely from what's in your project — no external sources, no drift.
Deep Research
For substantial research tasks, Deep Research runs multiple searches in sequence, synthesizes across sources, resolves contradictions, and delivers a structured report. It takes a few minutes and is significantly more thorough than asking a single question.
WHEN TO USE DEEP RESEARCH
Researching a new prospect or industry before a pitch
→
→Benchmarking your pricing or positioning against competitors
→Understanding a regulation, legal development, or compliance requirement
Building a market brief for a new service you're considering
→
Compiling a competitive landscape for a customer
→
I need a competitive landscape for [INDUSTRY/NICHE] in [REGION].
Cover: the 5 main players, how they position, their pricing if available, what they're doing well and where they're weak. Include any notable industry shifts or regulatory changes in the last 12 months.
Structure it as an executive brief I can share with my team.
Credit Usage Optimization
Claude doesn't count messages — it measures tokens. Every word you send, every word Claude returns, and every file or document loaded into context costs tokens. Understanding how this works means you stop hitting limits at the wrong moment and get significantly more out of every dollar you pay Anthropic.
How the limit system actually works
Usage is tracked across a 5-hour rolling window. It's not a daily cap that resets at midnight — it's a moving window. When your 5-hour total hits the threshold, Claude pauses. As old usage ages out of that window, capacity returns. There's also a weekly cap that most users don't discover until they've been working heavily for several days and suddenly have nothing left on a Thursday afternoon. The weekly cap now resets at a fixed day and time assigned to your account — check Settings → Usage to see exactly when yours comes back.
Claude.ai, Cowork, and Claude Code all pull from the same pool. A heavy morning session in Cowork leaves you with less for afternoon chat work. Plan your high-volume tasks accordingly. And note that Fable 5.1 — the top-tier model — is the exception to the pool: on Pro and Team Standard seats it isn't covered by your plan at all and runs on pay-as-you-go usage credits from the first message.
What each plan actually gives you
| PLAN | 5-HOUR WINDOW | WEEKLY CAP | PEAK THROTTLING |
|---|---|---|---|
| Pro $20/mo | Baseline. Enough for focused | Yes — resets at a fixed | Previously throttled |
| daily work — hits limits on heavy | weekly time assigned to | weekday mornings (5– | |
| document processing or long | your account (see | 11am PT). Eliminated as | |
| Cowork sessions. | Settings → Usage) | of May 2026. | |
| Team | 1.25× Pro per seat. Each seat | Yes — same fixed weekly | Eliminated as of May |
| Standard | has its own independent pool — | reset as Pro, across all | 2026. |
| $25/seat | one heavy user doesn't affect anyone else. | models | |
| PLAN | 5-HOUR WINDOW | WEEKLY CAP | PEAK THROTTLING |
| Team | 6.25× Pro per seat. Built for | One weekly cap across | Eliminated as of May |
| Premium | power users running Claude | all models. Fable 5.1 is | 2026. |
| $125/seat | Code, long Cowork sessions, or | included, up to half of the | |
| large document batches daily. | weekly allowance |
Team plan: you can mix seat types, and the minimum is now just 2 seats. Most people on a team never hit ℹ the Standard cap — marketers, project managers, writers doing daily communication work are fine on Standard. Assign Premium seats only to people who regularly do heavy batch work, large document processing, or long Cowork automations. Admins can also purchase overage credits so no one gets blocked mid-task.
The highest-impact ways to reduce usage
1 — USE SONNET FOR MOST WORK, OPUS ONLY WHEN IT MATTERS, FABLE ALMOST NEVER Opus burns through your allowance faster than Sonnet, and Fable 5.1 faster still — on Pro it isn't in your plan at all and bills separately as usage credits. For emails, proposals, summaries, and customer communication — the work that makes up 90% of a small business owner's Claude use — Sonnet 5 is the right model, and it's now close enough to Opus that you rarely lose anything. Reserve Opus 5 for the tasks where depth of reasoning genuinely changes the output: contract review, strategic analysis, high- stakes writing.
2 — BATCH RELATED QUESTIONS INSTEAD OF ASKING SEQUENTIALLY Every new message you send includes the growing conversation history as context — it costs tokens each time. If you have three related questions, send them together in one structured prompt. You get the same answers and spend roughly a third of the tokens.
TOKEN-HEAVY Message 1: "Summarize this email thread." Message 2: "Now draft a reply."
"Summarize this email thread, then draft a short reply (3 sentences max) that [specific goal]." One round trip. Same result.
Message 3: "Make it shorter." Three round trips. The full thread is loaded into context three times.
3 — KEEP YOUR SYSTEM PROMPT LEAN Your project system prompt loads on every single message. Every word in it costs tokens every time. Keep it under 200–300 words where possible. Cut anything that doesn't change how Claude behaves — vague philosophy, redundant statements, padded introductions. Dense and specific beats long and thorough.
4 — START FRESH CONVERSATIONS FOR LARGE TASKS As a conversation grows, every message re-sends the accumulated context. A 30-message thread where you're processing a long document costs dramatically more per exchange than it did at message 5. For batch processing, large file analysis, or any multi-step automation — start a new conversation rather than stacking onto an existing one.
5 — USE CONNECTORS INSTEAD OF PASTING DOCUMENTS Pasting a full document into a prompt loads its entire content as tokens. When you ask Claude to reference a file via a connector (Drive, Notion), Claude fetches only what's relevant rather than consuming the entire document. For large documents you reference regularly, connecting the source tool is more efficient than pasting every time.
6 — SCHEDULE HEAVY COWORK TASKS OUTSIDE PEAK HOURS If you're running large batch Cowork automations, plan them for evenings or weekends rather than your peak working hours. This preserves your 5-hour window for interactive work during the day when you need it most.
7 — TURN OFF WEB SEARCH WHEN YOU DON'T NEED IT Web search adds tokens for every search result Claude reads before answering. If you're working from your project context — drafting, editing, prompting against your own data — toggle the globe icon off. It's
a small saving per message, but across a full working day it adds up.
ON TEAM PLANS: THE SEAT YOU ACTUALLY NEED Before upgrading a seat to Premium ($125/seat), ask: does this person regularly hit their Standard cap? Most don't. The use cases that genuinely need Premium are: running Claude Code for hours daily, processing large document batches repeatedly, or heavy Cowork automation combined with active chat use. Everyone else — Standard.
The single highest-leverage habit: Apply the batch principle (tip 2) and the model selection principle (tip 1)
consistently. Together, independent testing suggests they reduce token consumption by 40–50% for typical knowledge-worker usage. That's the difference between hitting your limit at 2pm or never hitting it at all on most days.
Cowork
Cowork is Claude operating autonomously — reading, editing, and creating files, using your apps and browser, and running multi-step jobs without you watching every step. It started on the desktop app and now runs on the web and your phone too. This is Claude as an AI operator, not a chatbot.
ℹCowork comes in two flavours now. Local Cowork runs in the Claude Desktop app (Mac and Windows, generally available since April 2026) and works inside a folder on your computer that you designate — everything else on your machine is off-limits unless you switch on computer use. Cloud Cowork (July 2026, rolling out from Max plans to the rest) runs your session on Anthropic's side, so it's available from the web and the mobile app, your files and sessions follow you across devices, and work keeps going after you close your laptop. Included on every paid plan.
What Cowork can do that Claude Chat cannot
| CAPABILITY | CHAT | COWORK |
|---|---|---|
| Read files on your computer | — | ✓ |
| Edit and save files directly | — | ✓ |
| Create multiple files in one task | — | ✓ |
| Run scheduled tasks automatically | — | ✓ |
| Work while you're away from your desk | — | ✓ |
| Click, type, and navigate your apps and browser (computer use) | — | ✓ (opt-in) |
| Assign tasks from your phone and check back later Use Skills | — | ✓ |
| ✓ | ✓ | |
| Connect to third-party MCP servers | ✓ | ✓ (plus local ones) |
The practical payoff for small business owners
The shift from Chat to Cowork isn't a feature upgrade — it's a different kind of work delegation. In Chat, you give Claude one task at a time and wait. In Cowork, you describe a multi-step job and Claude executes it across your actual files while you do something else.
Process 30 PDFs in a folder, extract customer names and project values, build a summary spreadsheet — while
→
you're in a meeting Read every contract in your Customers folder, flag anything expiring in 90 days, create a renewal alert document
→
Rename every file in a folder to match a naming convention you define once
→
→Generate a weekly summary report from everything that changed in a folder — automatically, every Monday
How to get started with Cowork
1 Open Cowork — desktop app or the Cowork tab on claude.ai
For work on files that live on your computer, download the desktop app for Mac or Windows from claude.ai. For cloud sessions, open the Cowork tab in your browser or the mobile app. Either way, sign in with your existing account — your projects, system prompt, and connectors carry over; chat and Cowork now share one home.
Point it at a working folder (local) or your connectors (cloud)
2
On the desktop app, select a folder on your computer. Keep it focused — a "Claude Work" folder that you put things in deliberately, not your whole desktop. Cloud sessions work off your connected tools (Drive, Gmail, Notion) and files you upload instead.
3 Give Claude a multi-step task in plain language "Go through all the proposal PDFs in the Projects folder, extract each customer name and total value, and create 4 Review the output and refine Cowork shows you a log of what it did. Check the first few outputs carefully before relying on them for anything
Back up your folder before starting any new Cowork task. Claude can edit and overwrite files. Until you know how it behaves on your setup, keep backups of anything you'd miss.
Skills
A Skill is a reusable workflow you write once in plain English, then reuse forever — by name, with a slash command, or just by asking for the task it covers. The difference between a prompt you type every time and a skill you store is the difference between work and overhead.
How Skills work
A Skill is a folder with a plain-text file (SKILL.md) that describes a process in detail — what to do, in what order, with what output format — plus any templates or reference files it needs. Claude reads those instructions automatically when the task matches, or when you call the skill by name. No coding, no formulas. If you can describe a process to a capable colleague, you can write a Skill. Anthropic's own document skills (Word, Excel, PowerPoint, PDF) already run behind the scenes every time Claude builds you a file.
Skills now work everywhere — claude.ai in the browser, the desktop app, Cowork, and the Excel and PowerPoint add-ins. They're managed under Customize → Skills (Settings → Capabilities on some plans) and need "Code execution and file creation" switched on. Think of them as your personal library of encoded workflows.
Example Skills for everyday business work
SKILL /proposal /followup /weekly SKILL /scope /brief /invoice- review
WHAT IT DOES WHEN CALLED Generates a complete proposal from a brief — your standard structure, your pricing language, your tone, no blanks Drafts a post-meeting email summarizing what was discussed, decisions made, and next steps with owners Reads your calendar, Gmail, and open tasks; produces a structured weekly priority summary WHAT IT DOES WHEN CALLED Converts rough notes from a customer call into a clean, structured scope of work document Turns a customer intake form or email thread into a formatted internal creative/project brief Pulls AR aging from QuickBooks, flags what's overdue, drafts collection follow-ups for each
How to create a Skill
1 Open Customize → Skills On claude.ai or in the desktop app, open Customize (or Settings → Capabilities) and find the Skills section. 2 Have Claude build the skill for you
Turn on the built-in skill-creator skill and say "create a skill called proposal." Claude interviews you about the process, writes the SKILL.md, and hands you a ZIP to upload — no manual file editing. Describe every step Claude should follow, the format of the output, and any rules that apply.
| 3 Call it — or just ask for the task | ||
|---|---|---|
| In chat or Cowork, type after: /proposal for Meridian Co., website redesign, $8,500. Claude will also pick it up on its own when your request matches what the skill describes. | /proposal | or mention it by name and Claude loads the skill. You can add context |
Start with your single most-repeated task. The one you ask Claude to do at least once a week. Write it out in full detail once. After that, two words replace fifty. That's the ROI of a Skill — paid back in full by the end of the first week.
Scheduled Tasks
Scheduled Tasks let Claude run a workflow automatically at a time you set — every Monday morning, daily at 8am, the first of each month — without you initiating anything. You set it once and it runs. And since Cowork moved to the cloud, it can run even when your computer is off.
What you can automate
Monday 8am:Pull last week's Gmail activity and Drive changes, generate a weekly summary document
→
→Friday 4pm:Review open tasks and projects, draft a status update for your team
| →1st of the | Pull AR aging from QuickBooks, flag anything overdue 30+ days, draft one follow-up per late | |
|---|---|---|
| month: | invoice | |
| →Daily | Check today's calendar, pull any prep notes from Drive, build a morning brief with what needs | |
| 7am: →Weekly:Scan Gmail for unanswered threads older than 5 days, draft replies and save them as drafts | attention |
How to set up a Scheduled Task
Open Cowork → Scheduled Tasks
1
In Cowork (desktop app or the web), click Scheduled Tasks to see existing automations and create new ones — recurring or on-demand.
2 Write the task in plain English Describe exactly what you want Claude to do: "Every Monday at 8am, check my calendar for the week and my Set the schedule
3
Choose frequency (daily, weekly, monthly) and the exact time. Save it and verify it appears in the active task list.
Check the first few outputs
4
Find the output where you told Claude to put it. Review it the first two or three times to make sure the task is running correctly. Adjust the instruction if something is off.
Where the task runs decides whether your computer needs to be on. A task scheduled in a cloud Cowork session runs on Anthropic's side — no device needs to be online. A task scheduled in a local desktop session still needs Claude Desktop open and your computer awake, because it works on files on that machine. If you depend on scheduled automation and your plan has cloud Cowork, schedule there; if not, keep the desktop app running in the background.
Connecting Other Tools (MCP)
MCP is the open standard that lets Claude connect to external tools and read live data. Every connector — Gmail, Notion, QuickBooks, Slack — runs on MCP under the hood. You don't need to know how it works to use it, but knowing the mental model prevents confusion when things don't work as expected.
The mental model that matters
MCP is not a toggle. It's a live read. When Claude uses a connector, it sends a request to that tool in real time, gets data back, and pulls it into your current conversation's context window. The data isn't stored — it's fetched fresh each time and lives only for the duration of that conversation.
This means two things in practice:
| It's always → | When you ask Claude to check your Gmail, it reads today's emails — not a cached version | |
|---|---|---|
| current. | from yesterday. | |
| →It uses | The data Claude pulls occupies space in the conversation. If you pull in large documents or long | |
| context | email threads, that compresses the room for your conversation. For big retrieval tasks, start fresh | |
| budget. | conversations rather than stacking on top of long threads. |
Three types of MCP connection
🔗 Native Connectors Gmail, Drive, Calendar, Microsoft 365, QuickBooks, Xero, Notion, Slack, GitHub. Available directly in Claude Settings → Connectors. One-click to authorize. 🏠 Local MCP Tools running on your own machine — a local database, an internal app, a custom-built system. Requires technical setup but gives the deepest integration possible.
🔧 Third-Party MCP Any app that's published an MCP server. HubSpot, Figma, Airtable, Zapier (8,000+ apps). Browse the directory at claude.com/connectors or paste a server URL under Settings → Connectors — it works in chat and in Cowork. ⚡ Zapier Bridge If your tool isn't natively supported, Zapier's MCP bridge connects to 8,000+ apps. Build the automation in Zapier, get the MCP URL, paste it in — it works.
When you'd actually use this knowledge
→You use a CRM (HubSpot, Pipedrive) and want Claude to read your contacts and deal pipeline
→You use Airtable or Notion for customer management and want Claude to query specific tables
You want to connect a tool that isn't in the native connector list
→
You want to build a deeper connection to an internal system your team runs
→
For most day-to-day use, the native connectors cover everything. Come back to this section when you outgrow them.
Models
Claude runs on three everyday model tiers — Opus, Sonnet, and Haiku — trading off between depth and speed, plus a separate top tier above them for the hardest work. The default for most work is Sonnet. Here's when and why to change it.
| MODEL | BEST FOR | SPEED | NOTES |
|---|---|---|---|
| Opus 5 | Complex strategy, long documents, high- | Slower | Use when you'd want a senior |
| Deepest | stakes analysis, nuanced judgment calls | advisor reviewing something, not a | |
| where quality matters more than speed | fast answer. Anthropic pitches it as close to Fable's intelligence at half | ||
| Sonnet 5 | Everyday work — emails, proposals, | Fast | The right default for 90% of what |
| Balanced | summaries, research, customer | you do |
communication, most creative tasks. Now also strong on multi-step, agentic work (planning and using tools on its own)
| Haiku 4.5 | Quick lookups, short drafts, classification, | Fastest | Useful for high-volume, low-stakes |
|---|---|---|---|
| Fastest | simple transformations where you need a fast answer not a great one | tasks in Cowork automations |
Snapshot — current as of September 2026: the everyday lineup is Claude Opus 5 (July 2026), Claude Sonnet 5 (June 2026), and Claude Haiku 4.5. Sonnet 5 is the default on Pro — it replaced Sonnet 4.6, closed much of the quality gap with Opus while staying fast and inexpensive, and made real gains in multi-step, autonomous work. Opus 5 replaced Opus 4.8 a month later and narrowed the gap to the tier above it. Above all three sits Claude Fable 5.1, released September 1, 2026. Anthropic ships new versions, and occasionally whole new tiers, faster than any printed guide can keep up with — sometimes within days. The version numbers in your model dropdown are the source of truth for what's available today; this section teaches the logic for choosing between them, which doesn't change when the numbers do.
There's a tier above Opus, for the rare job that needs it — and it's billed differently. Claude Fable 5.1 (September 2026, succeeding Fable 5) is Anthropic's most capable generally available model, built for the hardest, most autonomous work — long agentic runs, dense document and spreadsheet work, multi-step research. It costs more and runs slower, so it's not for everyday use. Two things to know before you pick it from the dropdown. First, billing: on Max plans and Team Premium seats it's included, up to half your weekly allowance; on Pro and Team Standard seats it is not in your plan and runs on pay-as-you-go usage credits from the first message (a free-inclusion promotion for Fable 5 ended July 19, 2026, and never covered 5.1). Second, safeguards: if a request touches certain sensitive areas, Fable quietly hands that turn to Opus — you'll see a note in the conversation. Worth remembering, too, that this tier went offline for three weeks in June 2026 under a government directive before being restored on July 1 — the newest, most powerful model is also the one most likely to see access changes. For day-to-day business work, Opus, Sonnet, and Haiku are the three that matter; only reach for Fable on the rare task where nothing else will do.
How to switch models
In claude.ai, click the model name shown at the top of the chat input field. A dropdown appears. Select the model you want. The switch applies to the current conversation only — your default stays the same.
You don't always have to choose. Claude increasingly routes work automatically — sending simple steps ℹ to a faster model and harder ones to a deeper model behind the scenes (this is most visible in Cowork and Claude Code). Manual selection still matters when you want to force quality up (Opus for a high-stakes task) or force speed up (Haiku for bulk processing). For everything else, leaving it on the default is fine.
Practical model selection for small business owners
→Drafting a proposal for a customer you care about:Sonnet. Fast enough, quality is there.
→Reviewing a complex contract for risk:Opus 5. Use the better judgment.
Processing 50 emails to extract customer names:Haiku in a Cowork automation. Speed is the value.
→
Writing a strategy memo for a major pitch:Opus 5. You're not in a hurry and quality matters.
→
→A one-off, genuinely hard job — say, reconciling a year ofFable 5.1, on a Max or Premium seat. Expect it
messy spreadsheets into a clean model: Everything else in daily operation:Sonnet 5. Leave it there. →
to draw down your allowance fast.
Plans
Claude has four paid tiers — Pro, Max, Team, and Enterprise. Here's what each includes and when to move up.
| PLAN | PRICE | WHO IT'S FOR | KEY FEATURES |
|---|---|---|---|
| Pro Solo | $20/mo | Solo users who don't | All everyday models (Sonnet 5 default), |
| need shared projects or | web search and Research, Projects, | ||
| admin controls | Artifacts, Connectors, Cowork, Claude |
Code, Claude Design, Skills, Scheduled Tasks. Fable 5.1 available on usage credits only
| Max | $100/mo (5×) or | Solo heavy users — daily | Everything in Pro + 5× or 20× the usage, |
|---|---|---|---|
| Solo power | $200/mo (20×) | long sessions, big | Fable 5.1 included (up to half your weekly |
| documents, Claude Code | allowance), first in line for new features like | ||
| or Cowork automation | cloud Cowork | ||
| Team | $25/seat/mo ($20 | Any team of 2 to 150 who | Everything in Pro + shared Projects, admin |
| Team | annual), min 2 | need shared projects and | dashboard, SSO, spend caps, 1.25× Pro |
| seats | admin controls | usage per seat, and the option to enable usage credits so nobody gets blocked mid- | |
| Team | $125/seat/mo | Heavy users on a Team | Everything in Team + 6.25× Pro usage per |
| Premium | ($100 annual) | plan — daily large- | seat (5× a standard seat) and Fable 5.1 |
| Power | document or automation | included. Mix freely: most of the team on | |
| work | standard, power users on Premium. | ||
| Enterprise | From | Larger organizations with | SSO, audit logs, custom roles, model |
| $20/seat/mo + | compliance requirements | controls, HIPAA option, custom data | |
| usage at API | — now self-serve, no | retention. Usage isn't bundled — it bills | |
| rates | sales call needed | separately, so heavy users cost more |
You pay Anthropic directly. Your subscription invoice comes from Anthropic, not Maz AI. Maz AI's fee was one-time, for the setup only.
Plans and pricing change. Tiers get added, prices move, usage limits get retuned. The structure above is current as of September 2026 — for the live picture, check claude.com/pricing.
When to upgrade
→Adding even one colleague who needs their own Claude access → Team plan (2-seat minimum now; for two
people it's often the better deal than two Pro seats)
→You want team members working inside shared projects → Team plan
You're hitting usage limits — Claude will tell you when this happens → Max (solo) or a Premium seat (Team)
→
You want to use Fable 5.1 regularly without paying per use → Max or Team Premium
→
You have SSO requirements, legal holds, or compliance audits → Enterprise
→
About usage limits
Claude Pro and Team have usage limits that reset on a rolling basis. Limits aren't measured in simple message counts — they're based on the size and complexity of what you're sending and receiving. Long documents, large file uploads, and multi-step Cowork tasks use more than short questions.
→If you regularly hit limits during heavy work days, Max (solo) or a Premium seat (Team) is the fix; Team admins
can also switch on usage credits Haiku and Sonnet use less budget than Opus — switching down for routine tasks extends your available usage
→
→Starting fresh conversations (rather than long threads) is more efficient — context build-up costs budget
Claude for Small Business is a product of Maz AI
Want this set up in your business?
The guide gives you the method. If you'd rather have it built into your business, the first conversation is free and takes twenty minutes.
Book a 20-minute call