GETTING STARTED GUIDE

Set up Claude the right way

A practical walkthrough for getting your team productive with Claude — from the first login to building reusable workflows. Aligned with the V.E.P. framework from our AI Demystified series.

The V.E.P. Framework
VVerify Check facts, dates, numbers, citations
EEdit Replace generic with your voice
PPersonalize Add the human context only you have

① The V.E.P. Mindset

Before any setting, prompt, or feature — adopt the right mental model. Claude is a powerful starting point, not a source of truth.

What Claude actually does

Claude predicts the next word based on patterns from massive amounts of training data. Every response is a confident estimate — not a verified fact. That single shift in how you think about it changes everything that follows.

Key principle AI informs. The human decides. The "human in the loop" isn't optional — it's the design.

Where Claude shines

  • → Summarizing long documents, email threads, meeting notes
  • → Drafting first versions of emails, reports, SOPs, proposals
  • → Finding patterns in spreadsheets and datasets
  • → Brainstorming options and pressure-testing decisions
  • → Acting as a critical reviewer of your own draft

Where Claude falls short

  • → Precise math and calculations — always verify
  • → Recent events outside its training data
  • → Specific names, dates, citations, regulatory details
  • → Understanding why a document matters to your client
  • → Reading the room, judgment calls, relationship context

The V.E.P. framework — apply to every output

V — Verify

If important information is involved, check the date, the name, the number, the citation. Claude is statistically likely to be wrong on those because it learned from older data and predicts what sounds right.

E — Edit

Generic out of the box. Replace boilerplate with your voice, your company's tone, your domain knowledge. The first draft is a starting point, not the deliverable.

P — Personalize

Add the context only a human with your relationship and history can add. This is where you create the white-glove difference — the human touch that AI cannot replicate.

Why this matters When everyone on your team applies the same three-step review, output quality stops depending on who ran the prompt. You get consistency — almost a brand standard for how your organization uses AI.
Watch out Hallucinations aren't a bug — they're a function of how prediction works. The most common failure points: citation numbers, regulatory details, specific names, and dates. Slow down on those.

② Organization Setup

For Team and Enterprise plan owners. Get this right once and every user inherits the foundation. Skip ahead to Personal Settings if you're an individual user.

1. Set organization instructions

Organization instructions apply to every conversation across your account. They're the place to put company-wide standards: tone, formatting rules, compliance reminders, and domain context.

  1. Sign in as Owner or Admin → click your initials (lower left) → Settings
  2. Open Organization settings → Organization and access
  3. Find the Organization instructions section
  4. Enter your instructions and save

How to draft yours — let AI help

Don't stare at a blank box. Use Claude (or any AI tool) to draft your first version. Paste this prompt:

You are helping me draft "organization instructions" for our company's Claude account. These will apply to every conversation every employee has with Claude. About my organization: - Industry: [your industry] - What we do: [one or two sentences] - Audience our team writes for: [clients / internal / regulators / mixed] - Compliance frameworks we operate under: [HIPAA, SOC 2, CMMC, none, etc.] - Our tone: [warm and direct / formal / technical / etc.] What I want Claude to consistently do: - [example: always lead with the answer, caveats after] - [example: avoid marketing fluff and hedging] - [example: search and verify before stating regulatory or product details] - [example: format reports in prose, not bullet lists, unless asked] What I want Claude to avoid: - [example: don't add generic legal/medical/financial disclaimers] - [example: don't expand short questions into long answers] Draft a clean, focused set of organization instructions (300-500 words max). Use plain language. Group related rules together. Output only the instructions — no preamble.

Generic example output

Here's roughly what a good set of org instructions looks like once filled in. Use as a reference, not a template — yours should reflect your business.

You are assisting employees of [Company Name], a [brief description]. Our clients are [audience]. Communication standards: - Be direct. Lead with the answer. Reasoning and caveats come after. - Match our voice: [warm peer-level / formal / technical / etc.]. No marketing fluff. - Default to prose. Use bullets, headers, or tables only for procedures, reference material, or content the user will reuse. Accuracy requirements: - For [regulated topic 1], [regulated topic 2], pricing, or product specifications: search and verify before answering. Say "I need to verify" rather than guess. - Never invent citations, control numbers, license SKUs, or contact details. - Verify any numerical claim before including it in client-facing material. When drafting client communication: - Use a direct opening — no "I hope this email finds you well." - Include specific evidence or data, not vague benefits language. - Always end with a clear next step, owner, and date. Don't: - Don't add "consult a professional" disclaimers — our employees are the professionals. - Don't soften pushback into mush. Stay on point.
Precedence Organization instructions take priority when they conflict with individual user preferences. Keep them universal — leave role-specific guidance to personal preferences.

2. Enable Skills (org-wide)

Skills are reusable capabilities you'll build in Section 7. Enable them at the org level first.

  1. Go to Organization settings → Skills
  2. Toggle on Code execution and file creation (required prerequisite)
  3. Toggle on Skills
  4. Toggle on Skill sharing so your team can share skills internally

3. Decide on data governance up front

Before rollout, document what your team is and isn't allowed to put into Claude. This is one of the most common gaps and a topic we cover in depth in our AI Governance & Security session.

Critical Free and personal AI accounts have far less protection than business plans. Treat anything you put into AI as potentially stored and reviewable. Sensitive or regulated data should never go into a free or personal account — and even on enterprise plans, you need a written policy on what's allowed.

At a minimum, your policy should answer:

  • → What categories of data are approved for upload? (general business, internal, confidential, regulated)
  • → Which data is prohibited? (client PII, PHI, payment data, credentials)
  • → Who approves exceptions and how is that documented?
  • → What happens if data is uploaded by mistake?

③ Personal Settings

Each user sets these. Org instructions cover company-wide rules; personal preferences cover who you are and how you work.

Personal Preferences
Response Styles

Set personal preferences

These apply to every chat you start. Five focused minutes here pays back every day after.

  1. Click your initials (lower left) → Settings → Profile
  2. Find Personal preferences
  3. Paste your preferences and save

How to draft yours — let AI help

Same approach as org instructions. Don't write from scratch. Paste this prompt into Claude:

Help me write personal preferences for my Claude account. These tell Claude who I am and how I want responses tailored to me. They apply to every chat I have. About me: - My role: [your title] - Industry / company type: [brief] - Who I write for in my work: [clients, internal team, leadership, vendors] - My experience level: [junior / mid / senior / expert] How I want answers: - [example: be direct, no "great question" preamble] - [example: lead with the answer, then explain] - [example: numbered steps for procedures, prose for analysis] - [example: include error handling in code, comment non-obvious logic only] - [example: match my tone — direct, conversational, no fluff] Things that annoy me about generic AI responses: - [example: long preambles before getting to the point] - [example: hedging and over-disclaimers] - [example: bullet lists when prose would be clearer] - [example: rewriting my voice out of edits] Topics I work in regularly (so Claude doesn't over-explain basics): - [list 3-6 topics, tools, or technologies] Draft a focused set of personal preferences for me, 300-500 words max. Use plain instructions. No formal headings — just clear directives. Output only the preferences.
Sweet spot 300-500 words. Long preferences burn context window before you've even asked a question. Refine over time — every time Claude does something annoying, add one line to fix it.

Generic example — knowledge worker

Role: I'm a [project manager / analyst / consultant / etc.] at [type of company]. I work primarily on [main areas of work]. Audience for my output: [clients / internal leadership / cross-functional team]. How I want answers: - Be direct. Skip filler like "great question" or long preambles. - Lead with the answer. Reasoning and caveats come after. - For procedures or step-by-step work: numbered lists with specific actions. - For analysis or recommendations: prose, not bullet salad. - Match my tone — professional but conversational. Not corporate. For draft documents and emails: - Direct opening. No "I hope this email finds you well." - Specific evidence over vague benefits language. - End with a clear next step. Verify before stating: - Numerical claims, percentages, dollar figures. - Specific dates, names, titles, citations. - Recent events or product specifications. Don't: - Don't add generic disclaimers ("consult a professional", etc.) — I am the professional. - Don't expand short questions into long answers. - Don't rewrite my voice when editing my drafts.

Response Styles

Styles control how Claude formats responses (tone, length, structure). Personal preferences control who you are. They stack — they don't override each other.

Switching styles

  1. In any chat, click "Search and tools" in the lower-left corner
  2. Click "Use style"
  3. Pick a preset, or click "Create & edit styles" to build your own

The four presets

StyleWhen to use it
NormalGeneral questions, mixed work — the default
ConciseQuick answers, lookups, when you don't want to skim
FormalPolished business writing, official communications
ExplanatoryLearning a new topic, training, onboarding
Easiest win Switch your default to Concise for a week. Most people stop editing for length within a day or two and never go back.

Build a custom style from your writing

  1. "Use style" → "Create & edit styles" → "Create custom style"
  2. Paste 2-3 samples of writing in your voice (recent emails, a blog post, a report intro)
  3. Claude generates a style description — preview it
  4. Edit the description directly if it's off ("less formal", "shorter sentences")
  5. Save. Switch to it from the chat menu when relevant.

④ Prompting Workflow

If there's one thing to take away — learn to prompt well. The output is only as good as the input. This is the framework we teach in our AI Demystified series.

The four-step workflow

Combine these four moves into a single prompt for high-leverage work:

Step 1 — Role

Tell Claude who to be. "You are a senior financial analyst." "You are a skeptical reviewer." This single move noticeably changes the response.

Step 2 — Context (Data)

Give Claude what it needs to do the work — meeting notes, source documents, audience, tone, key points to cover. The more specific the input, the better the output.

Step 3 — Generate the draft

State the deliverable clearly. "Draft a one-page memo." "Write a client email." "Build an SOP."

Step 4 — Self-critique

This is the move most people miss. Ask Claude to review its own draft as a skeptical reviewer, find weaknesses, and produce an improved version. This is the V in VEP — built into the prompt.

Side-by-side: same task, different prompts

❌ Vague — generic output
Help me with the budget memo
✅ Full workflow — sharp output
You are a senior financial analyst. I need you to draft an executive budget memo for the CFO and leadership team. Here are the meeting notes from yesterday's planning session [paste notes].

Draft the first version including: executive summary, three key budget risks, recommended actions, and next steps with owners.

Then critique your own draft as a skeptical CFO. Identify the three biggest weaknesses. Fix them and provide only the improved final version.

Five core skills — practice these

1. Summarize

Probably the easiest place to start and the fastest "aha" moment. Long emails, reports, meeting notes, transcripts.

Summarize this email thread in five points. What were the decisions made? What's still open? What do I need to do to respond?
Pro tip Upload the file rather than pasting the text. Claude reads structured files (PDF, DOCX, XLSX, TXT) more efficiently — better answer, less token usage.

2. Role + Context

The foundation. Without these, you get generic. With them, you get tailored.

You are a senior account manager. Draft a response to [contact name], who is the VP of Operations at our largest client. Tone: warm but firm. Key points to cover: [list]. Audience expects directness, not corporate-speak.

3. Self-critique

The single highest-leverage move. Have Claude audit its own work before handing it back.

Act as a skeptical reviewer. Find the three biggest weaknesses in the draft above and explain why each is weak. Then rewrite the weakest section, matching the original tone.
Why this works Self-critique directly addresses hallucinations and sloppy reasoning. You're forcing Claude to slow down and pressure-test its own output before it lands on your desk. This is the V in VEP, baked into the prompt.

4. Brainstorm and pressure-test

Use Claude as a thinking partner — not just a writer. Get options, hear the counterargument, prepare for pushback.

Give me 10 ideas for [specific goal]. For each, one sentence on why it would work and one risk to watch out for.
I'm recommending [decision]. Argue against it. Give me the strongest case for the opposite. What would I need to change to win the argument?
I need to deliver difficult news to [person/role]. Draft my message — empathetic but direct. Then, as the recipient, predict how they'll push back. Revise the message to handle that pushback up front.

5. Iterate (repetitive prompting)

Research shows the third version of a prompt usually outperforms the first. Each follow-up holds more context. Don't accept the first answer as final — push back, refine, ask Claude to try again with a tweak.

Hallucination watch Slow down on: citation numbers, regulatory details, specific names and dates, recent events, precise calculations. These are exactly where Claude is most likely to be confidently wrong.

The combined workflow — single prompt

Once you're comfortable, combine all four steps into one prompt:

I'm uploading [a business scenario / source documents / meeting notes]. Your task — for [Company / Project / Client]: 1. ROLE: You are a [specific role]. Bring the perspective and standards of that role to everything you produce. 2. CONTEXT: Use the attached materials as the basis for your work. Don't make up facts not present in the documents. 3. DRAFT: Generate the following deliverable — [specific output: memo / email / SOP / proposal / etc.] Required sections: [list] Audience: [who will read it] Tone: [how it should sound] 4. SELF-CRITIQUE: Before giving me the final version, review your own draft as a skeptical [reviewer role]. Identify the three biggest weaknesses. Fix them. Provide ONLY the improved final version. After the deliverable, briefly note: what weaknesses you found in your first draft and how you fixed them.

⑤ Projects

A Project is a dedicated workspace with its own knowledge base, custom instructions, and chat history. The compounded value over time is enormous.

Why projects matter

A regular chat starts from scratch every time. A project carries your context, documents, and standards into every conversation. By month three, your project answers are dramatically better than week one — Claude knows your real materials, refined instructions, and the patterns that work for your team.

When to use a Project vs a regular chat

Use a Project forUse a regular chat for
Recurring work — anything you do more than twiceOne-off questions
Per-client work where the same context applies repeatedlyGeneral research
Initiatives with reference materials — SOPs, templates, brand guidesQuick lookups
Anything where the same documents need to be available across multiple chatsTasks where context changes every time
Best practice Scope each project narrowly — one project per client, or one per type of work (SOP drafting, compliance reviews, executive memos). The narrower the project, the better the output.

Create a project

  1. Sidebar → Projects → Create Project
  2. Give it a clear, descriptive name (you'll have many)
  3. Add a one-line description of what the project is for
  4. Visibility: keep private, or share with your team if you're on a Team plan

Three things to fill in for every project

1. Custom instructions

Click "Set project instructions". This is where you tell Claude the role, audience, voice, and rules for this specific work.

How to draft project instructions — let AI help

Help me write custom instructions for a Claude project. The project is for [type of work — e.g., "drafting client communications" or "preparing executive budget reviews"]. The work involves: - Audience: [who reads the output] - Cadence: [how often this work happens] - Documents typically referenced: [list types — SOPs, brand guides, prior reports, etc.] - Tone required: [warm, formal, technical, etc.] - Compliance or governance considerations: [any] Standing rules I want Claude to always follow in this project: - [example: always apply the V.E.P. framework — verify, edit, personalize] - [example: never include data not present in the uploaded files] - [example: ask before drafting if context is missing, don't guess] Draft the project instructions in 200-400 words. Output only the instructions text, ready to paste.

Generic example — project instructions

You are an AI assistant supporting [type of work] for [team / department / client name]. Your role: - Help draft, review, and refine [specific deliverables]. - Primary audience for outputs: [audience]. - Always apply the V.E.P. framework: verify facts, edit for our voice, personalize for the relationship. Writing style: - Warm, professional, direct. Match the tone of the reference documents in this project. - Lead with the recommendation or answer. Supporting detail follows. - Use prose for analysis. Use lists only for procedures, action items, or true reference content. Rules: - Use the uploaded reference documents as the source of truth. Don't invent facts not in the materials. - If important context is missing, ask before drafting — don't guess. - For numerical claims, dates, or names: flag them so the human can verify. - Treat all materials in this project as confidential. Default workflow when I ask for a draft: 1. Use the project's reference materials as context. 2. Generate the first version per the request. 3. Self-critique as a skeptical reviewer. 4. Provide only the improved version.

2. Reference documents (knowledge)

Drag files into the project. Claude will reference them in every chat within the project — no need to re-upload.

What to upload: templates, SOPs, brand guidelines, prior approved deliverables, contact lists, glossaries, anything Claude needs to "know" about this work.

Avoid Don't dump your entire SharePoint into a project. Targeted, well-organized documents produce better answers than data dumps. If a doc is over 100 pages, summarize it or split it into sections first.

3. Saved prompts

Once you find a prompt that consistently produces great output, save it inside the project. Anyone with access can reuse it. Build your library over time — every effective prompt becomes reusable for the whole team.

Suggested project structure

A sensible starting structure for most teams:

  • 📁 Internal — Standard Templates — your common templates and SOPs
  • 📁 Client — [Client Name] — one per active client engagement
  • 📁 [Department] — Recurring Work — e.g., "Sales — Proposals", "Ops — QBR Prep"
  • 📁 Personal — [Your Workstreams] — your individual recurring tasks

⑥ Memory

Memory lets Claude remember facts about you across conversations — your role, your style, your standards. Combined with Projects, it's how Claude gets sharper over time instead of starting from zero each chat.

The difference between memory and files

Memory storesFiles (in projects) store
Facts about you — your role, your team, your writing style, your preferences Documents Claude can reference — SOPs, templates, reports, source materials
Carries across all chats automatically Available within the project where they live
Updates as you work — learns from what you tell it Static reference — you upload, Claude reads

Enable memory

  1. Click your initials → Settings → Capabilities
  2. Toggle on "Generate memory from chat history"
  3. Toggle on "Search and reference past chats"

Manage what Claude remembers

  • → To update memory mid-chat: just tell Claude — "remember that I work at X now" or "forget that I'm working on Y project"
  • → To review memories: ask "what do you remember about me?" — Claude will show you
  • → To start fresh without affecting memory: use Incognito Conversations from the chat menu

Power move: an explicit "How I Work" file

Beyond memory, advanced users create a single uploaded file inside a project called something like "How I Work." It captures detailed rules — how to handle emails, how to draft executive material, what governance to follow, your VEP standards. Then in any chat, you say:

Reference the "How I Work" file in this project for how I write and the rules you should follow. Then [do the task].

Template — "How I Work" file

# How I Work This file describes how I work and the rules Claude should follow when drafting on my behalf. ## My role and audience [Your role, who you work for, who reads your output] ## My voice - [3-5 specific traits — direct, warm, technical, no marketing-speak, etc.] - [Things you avoid — passive voice, hedging, etc.] ## When drafting emails - Direct opening, no "I hope this email finds you well." - One-line context, the point, evidence (max 3 bullets if needed), clear next step. - Match the recipient — peer-level for known contacts, more formal for first contact. ## When drafting executive materials (memos, reports, summaries) - Lead with the recommendation. Detail follows. - Include data and verifiable evidence — flag any number I need to verify before sending. - Use prose. Lists only for action items or true reference content. ## Governance rules I always follow - Apply V.E.P.: verify facts, edit for voice, personalize for relationship. - Never include facts not present in source materials. - Flag anything that needs human verification before it goes out. ## When I ask for help - Default to the four-step workflow: role → context → draft → self-critique. - Skip the self-critique only if I explicitly say "no review needed."
Confidentiality Memory is per-user, not per-organization. Be intentional about what you tell Claude to remember — especially around sensitive client matters. For confidential work, scope it to a Project where context stays bounded.

⑦ Skills

Skills are reusable instruction packages Claude loads automatically when relevant. Think of them as a "prompt within a prompt" that captures your brand identity, standards, or specialized workflows — and triggers itself when you need it.

Why skills matter

Projects scope Claude to one workstream. Skills go further — they carry your brand identity, voice, or specialized workflow into every chat where they're relevant. When a skill is active, Claude looks at your request through that lens automatically. Defined skills push generic AI output toward your organization's specific perspective without you having to re-prompt every time.

How skills work

A skill is a small package — at minimum, a single text file with metadata describing when to use it. Claude reads the metadata always; loads the full instructions when your request matches. You don't have to invoke them by name — Claude triggers them based on what you ask.

Enable skills in your account

  1. Settings → Capabilities → enable Code execution and file creation
  2. Customize → Skills → toggle on the example skills you want (Word doc, Excel, PowerPoint, PDF are common starting points)
  3. Organization-provisioned skills appear automatically once enabled at the org level

What to build first

Don't try to build a skill library on day one. Start with one skill that captures your most repeated request:

Skill typeTriggers when…Best for
Brand voiceDrafting any external contentClient-facing teams
Email toneDrafting emails to clients or vendorsAnyone with external comms
Resolution notesWrapping up tickets or service workService / support teams
Executive summaryBuilding memos, briefs, exec updatesOperations, finance, leadership
Compliance summaryDrafting compliance updates or risk notesRegulated industries

How to build a skill — let AI help

You don't need to write skill files by hand. Tell Claude what you want and have it generate the skill for you:

Help me create a custom Claude skill for [purpose — e.g., "drafting client emails in our company voice" or "writing executive summaries with our standard structure"]. Context: - Who will use it: [me only / our whole team / department] - When it should trigger: [describe the situations Claude should activate this skill] - What good output looks like: [paste 1-2 examples of strong work in this category] - Rules to always follow: [list] - Things to never do: [list] Generate: 1. A SKILL.md file with proper frontmatter (name, description, when to trigger). 2. Make the description specific so Claude reliably triggers it — list trigger phrases and contexts explicitly. 3. The body should contain the rules, voice guidelines, and any structural requirements. 4. Include 1-2 examples in the file showing the desired output style. Output the complete SKILL.md file ready to package and upload.

Upload a custom skill

  1. Place your SKILL.md file in a folder named after the skill
  2. ZIP the folder — the ZIP must contain the folder itself, not just the file inside
  3. Customize → Skills → click + → Create skill → upload ZIP
  4. Toggle it on. It now triggers automatically when your prompts match its description.
  5. To share with your team: open the skill → Share → entire organization (requires the org owner to enable sharing)
Description matters most Skills only fire if Claude matches your prompt to the description. Be specific and slightly direct — list trigger phrases explicitly. Vague descriptions = skills that never trigger.
Security Never put credentials, API keys, or client secrets in a skill file. Skills are uploaded as files and visible to anyone you share with.

✓ Quick Start Checklist

A practical onboarding path. Click items to mark them done. The whole sequence takes a few hours spread across the first month.

Day 1 — 15 minutes

  • Sign in to Claude
  • Read Section 1 (VEP Mindset) — it changes how you'll use everything else
  • Use the prompt in Section 3 to have AI draft your personal preferences
  • Paste them into Settings → Profile → Personal Preferences
  • Enable Memory and Past Chats in Settings → Capabilities
  • Set your default style to Concise (Section 3)
  • Run 3 real prompts from your actual work — verify the voice and format

Week 1 — 30 minutes

  • If you're an org owner: complete Section 2 (org instructions, enable skills, document data governance)
  • Practice the four-step prompting workflow on 5 real tasks
  • Try the self-critique move on a draft you'd normally accept as-is — notice the difference
  • Create a Project for your most common type of recurring work
  • Add custom instructions to that project using the prompt template in Section 5
  • Upload 3-5 reference documents to the project

Month 1 — 1 hour spread out

  • Create projects for your other recurring workstreams (one per client or initiative)
  • Save 3 high-performing prompts inside their relevant projects
  • Build one custom Skill for your most repeated request
  • Refine your personal preferences — add one line for every annoyance you've noticed
  • Apply VEP to every output for a full week — make it muscle memory
  • Compare notes with one teammate using Claude — share what's working

Ongoing habits

  • Every time Claude does something annoying — add a line to preferences to fix it
  • Every time you re-type the same prompt twice — save it to a project or convert it to a skill
  • Every output that goes external gets the full VEP review — verify, edit, personalize
  • Quarterly: prune unused projects and skills to keep things focused
Continue learning This is the foundation. Our AI Demystified series goes deeper on data and your advantage, AI governance and security, and building agents and automations. Reach out to your Tech River contact to learn more.
Tech River · Getting Started with Claude
Aligned with the V.E.P. framework from our AI Demystified series
Questions? Contact your Tech River account team.