Standalone article · part of a sequenced guide
What you'll unlock: Configure once (preferences, toggles, privacy), ship in artifacts (not chat scrollback), and search with source rules — three habits that multiply everything else you learned in this playbook.
The Interface Like a Power User
Every feature of Claude.ai — the settings, the shortcuts, and the behaviours most users never discover
Chapter context
Your team lives in Claude.ai daily — but everyone uses default settings, copies from chat bubbles, and turns on web search for tasks that don't need it.This chapter turns the interface from a chat window into a configured workstation: standing instructions, artifact deliverables, and research you can defend in a meeting.
Is this chapter for you?
Do you open Claude.ai more than once per day for real work?
Yes — complete Concept 1 this week; preferences alone save hours monthly.
Do you produce documents, specs, or code Claude should reuse?
Yes — Concept 2 is mandatory; stop shipping from chat scrollback.
Do your questions depend on news, markets, or changing public facts?
Yes — Concept 3 with citation review; never single-pass for high stakes.
Does your org handle sensitive client or employee data in Claude?
Yes — read 1.7 and 1.8 before enabling search or shared artifacts.
Chapters 1–2 gave you the mental model and the economics. This chapter is where those ideas meet the Claude.ai interface — the layer most professionals stare at for hours but never configure deliberately.Power users don't discover more features by accident. They treat settings as leverage, artifacts as deliverables, and web search as a directed research instrument — not a magic 'make it current' button.
Chapter insight
Configure once (preferences, toggles, privacy), ship in artifacts (not chat scrollback), and search with source rules — three habits that multiply everything else you learned in this playbook.
Reference diagrams
Claude.ai power-user stack
Settings shape every session; artifacts hold deliverables; search extends reach beyond your uploads.
Artifact workflow loop
Chat negotiates; artifact ships. Export closes the loop.
Implementation paths
Three concepts — configure, build, research.
Concept 1
Settings Deep Dive
The configuration layer that most users ignore — and the leverage it creates
1.1
User preferences
How to set your default communication style, format, and response length so you stop repeating yourself in every conversation
Key takeaway
Preferences are your standing brief to Claude — set tone, format, and brevity once; every new chat inherits them without a preamble.
Why this matters
Power users who re-type 'be concise, use bullets, no fluff' fifty times a week are paying a tax preferences eliminate.
Claude.ai user preferences let you specify how Claude should communicate: direct vs exploratory, formal vs casual, default list vs prose format, and how much detail to include.
Treat preferences like a job description for your assistant: role ('you support a VP of Product'), output norms ('lead with recommendation, then rationale'), and anti-patterns ('never open with Certainly!').
Keep preferences stable and generic. Task-specific constraints belong in the message or Project — not in global prefs.
Workflow — do this next
- 01Open Settings → Profile / Preferences.
- 02Write 3 lines: role, format default, length default.
- 03Start 3 test chats with one-word prompts — verify style holds without reminders.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
User preferences starter template
ROLE: I am a [role] at [type of org]. FORMAT: Default to [bullets / tables / short paragraphs]. Lead with the answer; put caveats after. LENGTH: [Concise / Standard / Deep dive when I ask]. Skip preamble and filler phrases. AVOID: [List 2–3 pet peeves, e.g. hedge words, unsolicited emojis].
1.2
Custom instructions
The persistent context that follows you into every new conversation — what to put here and what not to
Key takeaway
Custom instructions are global memory for how you work — not a filing cabinet for project docs.
Why this matters
Stuffing 2,000 words of company wiki into custom instructions burns context on every chat and dilutes what actually matters.
Custom instructions travel with your account into new conversations. Ideal content: standing constraints, vocabulary, and workflow habits.
Put here: industry jargon definitions, approval chains ('flag legal if customer data'), preferred citation style, tools you use daily.
Do not put here: full product specs, client decks, or anything that changes weekly — use Projects instead.
Workflow — do this next
- 01Audit current custom instructions — highlight anything project-specific.
- 02Move project content to a Project; keep <400 words global.
- 03Add one line: 'If unsure which doc applies, ask before assuming.'
Real example
Founder — global vs Project split
A founder kept investor update tone in custom instructions and moved each fundraise's metrics into a Project. New chats felt personal; fundraise work stayed scoped and versioned.
1.3
Memory settings
What Claude's memory system stores, how to view it, how to edit it, and when to clear it
Key takeaway
Memory is Claude's notebook about you — useful for continuity, dangerous if wrong facts stick.
Why this matters
Unreviewed memory causes confident wrong callbacks ('you prefer Opus for everything') that poison future sessions.
When enabled, Claude may retain memories across chats. You can view, edit, and delete individual memories in settings.
Good memories: your role, standing preferences you confirmed, long-running initiatives. Bad memories: one-off tasks, stale decisions, guesses you never verified.
Clear memory when changing jobs, after a major pivot, or if Claude keeps citing outdated context. Pair memory with Projects for anything team-visible.
Workflow — do this next
- 01Settings → Memory — read full list monthly.
- 02Delete anything you wouldn't want in a new hire's onboarding doc.
- 03After a wrong callback, fix or delete the source memory immediately.
1.4
Feature toggles
Web search, code execution, file creation, artifacts, extended thinking, Skills — what each does and when to turn them on or off
Key takeaway
Toggles are capability switches — artifacts ON for deliverables; search ON for research threads; code execution ON for analysis/skills with scripts; thinking ON only for hard tasks; Skills enabled per governance policy.
Why this matters
Web search on every chat invites stale citations; code execution on sensitive docs raises exfiltration risk; thinking always-on burns limits.
Key toggles (availability varies by plan): web search, artifacts, analysis/code execution, file creation, extended thinking, Agent Skills. Each adds tools, tokens, or execution surface.
Default stance: artifacts ON for build work; search ON only for research; code execution ON when verified runs or script-backed skills needed; thinking OFF until thread marked HARD; skills from trusted sources only.
Workflow — do this next
- 01List your top 5 task types.
- 02For each, note ideal toggle state.
- 03Create a pinned note in your Project: TOGGLE_DEFAULTS.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Feature toggle matrix
TASK TYPE | SEARCH | ARTIFACTS | CODE RUN | THINKING | SKILLS Drafting / editing | OFF | optional | OFF | OFF | if needed Live market scan | ON | ON | OFF | OFF | research skill Data / spreadsheet | OFF | ON | ON | OFF | xlsx skill Hard strategy/ADR | OFF | ON | OFF | ON | optional Build mini-tool | OFF | ON (React)| ON | OFF | OFF Sensitive internal | OFF | ON | OFF | OFF | internal only
1.5
Notification and interface preferences
The settings that make Claude faster to reach and easier to use daily
Key takeaway
Friction in reaching Claude determines whether you use it or open another tab — optimise access, not just prompts.
Why this matters
Thirty seconds saved per open × 20 opens/day is hours per month.
Pin Claude (browser PWA or desktop app), set keyboard shortcuts where available, and tune notifications so you only get interrupted for async jobs that matter.
Dark mode, font size, and sidebar defaults reduce fatigue on long sessions. Mobile vs desktop: reserve phone for capture and quick Q&A; deep artifact work on desktop.
Workflow — do this next
- 01Install desktop or PWA if you use Claude daily.
- 02Disable non-essential notifications.
- 03Bookmark your primary Project — not just claude.ai/new.
1.6
API key management for Claude.ai
Connecting your own API key and what that unlocks
Key takeaway
Bring-your-own-key (where offered) shifts spend to your API console — useful for power users who already meter production usage.
Why this matters
Teams mixing personal Pro with company API keys need one owner for keys, rotation, and spend alerts.
Some Claude.ai flows let you attach an Anthropic API key for higher caps or specific features. Keys live in console.anthropic.com — never in shared docs.
Unlocks: predictable API pricing for heavy users, alignment with org key management. Risks: accidental spend without console alerts; key leakage if pasted in chats.
Workflow — do this next
- 01Create a dedicated key named 'claude-ai-personal' in console.
- 02Set monthly spend cap and email alert.
- 03Rotate if ever pasted into a thread or screenshot.
1.7
Data and privacy settings
What Anthropic stores, what you can delete, and the controls that matter for sensitive work
Key takeaway
Know what persists — chat history, memories, uploads — and delete on schedule for sensitive work.
Why this matters
Compliance incidents are usually 'we didn't know it was stored there,' not model malice.
Review data retention and training policies for your plan (consumer vs commercial differ). Use chat deletion and memory controls for client data. Do not upload what your policy forbids — settings cannot un-see a file.
For regulated work: Enterprise/Team policies, SSO, and documented data handling beat consumer toggles.
Workflow — do this next
- 01Read current Anthropic privacy docs for your tier.
- 02Define team rule: which data classes may enter Claude.ai.
- 03Monthly: purge completed sensitive threads.
1.8
Account and billing management
The controls that prevent billing surprises and plan management mistakes
Key takeaway
Billing settings are where plan tier, payment method, and seat changes actually happen — check after every org change.
Why this matters
Shadow Pro seats and forgotten annual renewals show up here, not in usage limits.
Centralise payment for teams; avoid five cards for five PMs. Align Claude.ai subscription with API console billing owners from Chapter 2.
Before downgrade: export critical Projects. Before upgrade: confirm which friction you're solving (limits vs collaboration vs compliance).
Workflow — do this next
- 01Calendar quarterly billing review with finance.
- 02Document plan owner and escalation for limit hits.
- 03After team changes, reconcile seats within 48 hours.
Concept 2
The Artifact System
Claude's output layer — how artifacts work, what they enable, and how to use them as a workflow tool
2.1
What an artifact is
The separate output panel that makes Claude's work reusable rather than ephemeral
Key takeaway
An artifact is a durable object beside the chat — edit, copy, and export without losing it in message scrollback.
Why this matters
Chat-only workflows treat great outputs as disposable; artifacts turn responses into assets.
When Claude creates an artifact, it lives outside the message stream. You can iterate, select all, download, and sometimes share without re-prompting the entire thread.
Mental model: chat is the negotiation; artifact is the deliverable.
Workflow — do this next
- 01Next build task, ask: 'Put the final output in an artifact.'
- 02Name the artifact in your prompt (e.g. 'PRD artifact v1').
- 03Copy to your repo/Notion from the panel, not from chat markdown.
2.2
Artifact types
Code, markdown, HTML, React, SVG, Mermaid — what each is for and when Claude creates them
Key takeaway
Match artifact type to downstream tool — markdown for docs, code for repos, React for interactive prototypes.
Why this matters
Asking for React when you need a Word-ready memo wastes iteration cycles.
Common types: markdown, code files (syntax-highlighted), HTML, React, SVG/Mermaid for diagrams.
Prompt explicitly: 'Use a Mermaid artifact for the flow, markdown for the spec.' Claude picks type from task cues unless you override.
Workflow — do this next
- 01Define default type per task in your Project README.
- 02Request diagram type upfront for architecture work.
- 03For code, specify language and file extension in the prompt.
2.3
Editing artifacts
How to iterate on an artifact without restarting — the in-place editing workflow
Key takeaway
Iterate on the artifact object — 'edit section 3' or 'change the chart labels' — not the whole conversation from scratch.
Why this matters
Restarting threads loses constraints; artifact iteration preserves structure while refining content.
You can edit artifacts directly in the panel or ask Claude to modify specific sections. Reference the artifact by name or section heading.
Pattern: v1 artifact → review → 'In the artifact, tighten the executive summary and add a risks table' → v2 in same panel.
Workflow — do this next
- 01Review artifact before replying in chat.
- 02Give surgical edits, not 'try again'.
- 03Snapshot externally before major rewrites.
Real example
PM — spec iteration without thread rot
A PM kept a PRD in a markdown artifact through twelve revision rounds. Chat held decisions; artifact held the doc. Exec review used exported PDF from the artifact, not chat copy-paste.
2.4
Downloading and exporting artifacts
Getting Claude's work out of the interface and into your tools
Key takeaway
Export early and often — artifacts are not your system of record until they live in git, Drive, or your wiki.
Why this matters
Account limits, accidental deletion, and thread clutter make the interface a poor archive.
Use copy, download, or export actions on the artifact panel. For code, paste into your IDE or download as files. For markdown, push to Notion/GitHub with your normal workflow.
Establish a naming convention: YYYY-MM-DD_artifact-name_v3.md
Workflow — do this next
- 01After each approved version, export to canonical storage.
- 02Link exported file in Project instructions.
- 03Never treat Claude as sole backup for client deliverables.
2.5
Artifacts with persistent storage
How artifacts can store and retrieve data across sessions — the feature most users haven't found
Key takeaway
Some artifact flows support state across turns or sessions — treat as lightweight app memory, not a database.
Why this matters
Trackers and dashboards fail when users don't realise data must be re-loaded or saved explicitly.
Interactive artifacts (especially React) can hold in-session state. Cross-session persistence depends on product behaviour — verify what's saved when you return tomorrow.
For durable data: export JSON/CSV from the artifact or sync to your tools via MCP/API — don't rely on chat persistence alone.
Workflow — do this next
- 01Test: close tab, reopen — what survives?
- 02Document persistence limits in your team wiki.
- 03Use Projects for data that must survive weeks.
2.6
React artifacts as mini-applications
Building interactive tools inside Claude — calculators, dashboards, trackers
Key takeaway
React artifacts are prototypes in minutes — calculators, scenario models, approval checklists with UI.
Why this matters
Stakeholders understand a clickable model faster than a wall of text.
Ask Claude to build a React artifact with explicit inputs, validation, and labels. Great for finance scenarios, prioritisation matrices, and training simulators.
Limits: not for production auth, PII, or scale. Export logic to your codebase when validated.
Workflow — do this next
- 01Specify inputs, outputs, and one example calculation.
- 02Request 'all assumptions visible in UI'.
- 03Screenshot + export code when stakeholders sign off.
Real example
Ops lead — pricing scenario calculator
Built a React artifact with three levers (volume, discount, churn). Leadership changed assumptions live in a meeting — no spreadsheet email thread.
2.8
The artifact workflow mindset
Using artifacts as deliverables, not just outputs — the shift that changes how you assign work to Claude
Key takeaway
Assign Claude deliverables with artifact type, acceptance criteria, and export path — same as a human contractor.
Why this matters
Teams that say 'write me X' without artifact discipline get chat prose they can't ship.
Template every assignment: artifact type, structure, done-when checklist, export destination. Chat is for questions; artifact is what you ship.
Workflow — do this next
- 01Rewrite top 3 recurring tasks as artifact briefs.
- 02Add 'Definition of done' to each brief.
- 03Review exports in standup, not chat scrollback.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Artifact assignment brief
DELIVERABLE: [name] TYPE: [markdown | code | React | Mermaid] STRUCTURE: [sections / files] ACCEPTANCE: - [ ] Matches template - [ ] No placeholder text - [ ] Exported to [path] CHAT: Questions only — final work lives in artifact.
Concept 3
Web Search & Research Mode
Claude with eyes on the current world — how to use search effectively and what deep research actually does
3.1
How web search works inside Claude
What Claude searches, how it selects sources, and how it incorporates results
Key takeaway
Claude searches the live web, selects pages, reads excerpts, and synthesises — you get an answer with sources, not a single link dump.
Why this matters
Expecting Google-style exhaustive results leads to mistrust when Claude summarises selectively.
When enabled, Claude issues search queries, retrieves page content, and grounds its reply in what it fetched. Citations point to sources used — not every page on the internet.
Selection is model-judged: reputable docs, recent news, official sites — but mistakes happen on niche topics.
Workflow — do this next
- 01Ask one factual current-events question with search on.
- 02Click each citation — verify relevance.
- 03Note when citations are thin — signal to dig manually.
3.2
When to turn on web search
The task types that benefit from live information vs the ones where it adds noise
Key takeaway
Search ON for time-sensitive facts, prices, releases, regulations. Search OFF for drafting, coding, and strategy from your own docs.
Why this matters
Search adds latency and can pull irrelevant controversy into creative work.
ON: competitor moves, API changelog, market data, 'what happened this week', verifying a claim.
OFF: rewriting your memo, generating code from your repo context, roleplay, brainstorming from first principles.
Workflow — do this next
- 01Tag last 20 tasks: LIVE vs STATIC.
- 02Set default toggle from majority tag.
- 03Override per thread in first message when mixed.
3.3
Directing Claude's search
How to guide what Claude searches for rather than leaving it to default behaviour
Key takeaway
Give domains, date ranges, and forbidden sources — 'search SEC filings only' beats 'tell me about the company'.
Why this matters
Default search optimises for answer speed, not your evidentiary standard.
Direct with: specific entities, site: filters in plain language ('prefer .gov and primary sources'), recency ('since 2025'), and what to ignore (forums, SEO spam).
Workflow — do this next
- 01Prefix research prompts with SOURCE RULES block.
- 02Ask Claude to list queries it would run before searching.
- 03Reject answer if citations don't match rules — re-prompt.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Search directive block
SOURCE RULES: - Prefer: [official docs, .gov, primary research] - Avoid: [affiliate blogs, unsourced forums] - Recency: [last 12 months unless historical] - If insufficient sources, say "insufficient" — do not guess
3.4
Deep research mode
What it does differently from a single web search — the multi-source synthesis capability
Key takeaway
Deep research runs broader retrieval and synthesis — closer to an analyst deck than a quick fact check.
Why this matters
Users expect PhD-thesis rigor in 30 seconds; deep research needs time and still needs human review.
Deep research explores multiple angles, compares sources, and produces structured reports. Higher token and time cost than a single search turn.
Use for: market landscapes, policy comparisons, vendor shortlists. Not for: one-number lookups.
Workflow — do this next
- 01Define output schema before starting (sections, tables).
- 02Run deep research in a dedicated thread.
- 03Budget 15–30 min review time for citations.
3.5
Evaluating Claude's research quality
How to review sources, check citations, and catch hallucinations in research outputs
Key takeaway
Trust but verify — spot-check every high-stakes claim at its citation; one dead link is a flag to redo that section.
Why this matters
Confident wrong synthesis is worse than 'I don't know' — especially in board or legal contexts.
Review checklist: citation resolves, quote supports claim, date matches your recency rule, no circular sources (blogs citing blogs).
Hallucination pattern: specific numbers without citations, mismatched titles, 'according to reports' with no link.
Workflow — do this next
- 01Highlight uncited numbers — demand source or delete.
- 02Open 3 random citations per report.
- 03Second pass: 'challenge your three weakest claims'.
3.6
Research workflow design
Using Claude as a research assistant for ongoing work — the workflow that replaces hours of manual searching
Key takeaway
Standing research workflows: intake question → source rules → artifact report → human review → archive in Project.
Why this matters
Ad hoc search each Monday recreates prompts and inconsistent quality.
Weekly competitive intel: fixed template, search on, artifact markdown export to Project folder. Monthly: refresh source rules and retire stale memories.
Workflow — do this next
- 01Create RESEARCH_SOP.md in Project.
- 02Schedule recurring calendar block for review, not generation.
- 03Version reports with dates in artifact title.
Real example
Corp dev — weekly competitor monitor
Every Monday: same prompt shell, deep research for three rivals, markdown artifact to Notion. Analyst spends 25 min on citation QA, not 3 hours browsing.
3.7
Combining search with your own documents
The hybrid retrieval that uses both live web and uploaded context
Key takeaway
Best answers often blend your private docs (Project uploads) with live web — specify which wins on conflict.
Why this matters
Without precedence rules, Claude may prefer a random blog over your internal pricing doc.
Upload internal PDFs to a Project, enable search for external facts. Prompt: 'Internal docs override web on product facts; web for market size only.'
Workflow — do this next
- 01List facts: INTERNAL vs EXTERNAL.
- 02Upload internals; search for externals.
- 03State conflict resolution in first message.
3.8
Limitations of Claude's web search
What it misses, what it can't access, and how to supplement it
Key takeaway
Paywalled journals, logged-in portals, local PDFs, and real-time tickers often require human or MCP supplementation.
Why this matters
Shipping decisions on incomplete search is a process failure, not a model failure.
Gaps: paywalls, JS-heavy sites, non-English long tail, niche forums, anything behind your company SSO. Supplement with manual uploads, Claude in Chrome on allowed portals, or API pipelines.
When stakes are high, treat Claude research as first draft — expert review mandatory.
Workflow — do this next
- 01Maintain a 'search can't see this' list for your industry.
- 02Route those sources to Chrome extension or manual upload.
- 03Document gaps in research artifact appendix.
Concept 4
Claude.ai Capabilities — Complete Surface
Vision, extended thinking, Skills, code execution, model selection, sharing, and every major Claude.ai feature in depth
4.1
Vision workflows in Claude.ai
Uploading images and screenshots — UI review, diagram extraction, chart reading, and redacted incident photos
Key takeaway
Claude.ai accepts image uploads in chat and Projects — structure prompts for visible-only analysis; verify small text and numeric labels.
Why this matters
Vision is buried in 'attach file' — power users need explicit workflows to avoid hallucinated UI details.
Workflows: mockup → acceptance criteria; whiteboard photo → structured notes; dashboard screenshot → anomaly list (human confirms numbers). Always: 'If unreadable, say UNREADABLE.'
Workflow — do this next
- 01Crop to relevant UI region.
- 02Request table: element | observation | question.
- 03Designer/engineer validates before ticket creation.
Real example
Support — error dialog triage
User screenshot of error dialog uploaded. Claude extracted error code and visible stack fragment. Engineer matched to known issue in 5 min — full screen not needed after crop.
4.2
Extended thinking in Claude.ai
When to enable deeper reasoning in the consumer UI — latency trade-offs and task matching
Key takeaway
Toggle extended thinking for multi-step analysis, hard debugging, and strategic trade-offs — keep off for rewriting, formatting, and quick Q&A.
Why this matters
Thinking mode left always-on slows every interaction and burns usage limits.
Start thread with task complexity label. Enable thinking before hard question. Review thinking summary if shown — catch wrong assumptions early.
Workflow — do this next
- 01Default thinking OFF in preferences.
- 02Enable per-thread for HARD tasks.
- 03Compare output quality on one sample before habit change.
4.3
Agent Skills in Claude.ai
Browse, enable, upload, and govern skills — pre-built Office skills and custom zip uploads
Key takeaway
Claude.ai Skills: Customize → Skills → browse Anthropic/partner skills or upload custom zip (requires code execution on supported plans). Per-user — not org-centralized on consumer tiers.
Why this matters
Teams assume uploaded skills propagate — they don't without Team/API governance path.
Pre-built: Excel, PowerPoint, Word, PDF generation workflows. Partner: Notion, Figma, Atlassian skills in directory. Custom: SKILL.md + resources zipped; test in sandbox Project first.
Skills require code execution for script-backed skills — align with security policy (Ch 3.1.4).
Workflow — do this next
- 01Enable one pre-built document skill.
- 02Run golden test document.
- 03Upload custom skill; version in git.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Claude.ai skills governance
□ Skill source trusted (Anthropic / partner / internal)? □ Scripts reviewed for network/exfil risk? □ Test output on synthetic data first? □ Version recorded in team registry? □ Team members know per-user upload model?
4.4
Code execution & analysis mode
Python analysis in Claude.ai — spreadsheets, charts, data transforms, and security boundaries
Key takeaway
Code execution lets Claude run Python for verified math and charts — enable only when needed; never on raw secrets; export results after review.
Why this matters
Analysis mode is the bridge between hallucinated numbers and computed numbers.
Upload CSV → 'show code for each step' → charts in artifact. Disable on threads with credentials or unreleased financials if policy requires.
Workflow — do this next
- 01Sanitize CSV before upload.
- 02Require code visibility in output.
- 03Human spot-checks first row of results.
4.5
Model selection & conversation styles
Haiku, Sonnet, Opus routing in Claude.ai — speed vs depth, and style presets where available
Key takeaway
Pick model per task in UI: Haiku fast/cheap, Sonnet default workhorse, Opus hardest reasoning — don't default Opus for all work.
Why this matters
Ch 2 economics fail in practice when UI always stays on strongest model.
Match Ch 2 routing: extraction → Haiku; drafting → Sonnet; board-level strategy → Opus or Sonnet + thinking. Styles/presets (if available) encode tone — still verify substance.
Workflow — do this next
- 01Set Sonnet as default.
- 02Switch model in thread header for hard task.
- 03Log model used in exported artifact footer.
4.6
Sharing, starring & conversation management
Share links, organize chats, search history, and team visibility boundaries
Key takeaway
Share links export conversation snapshots — redact before sharing; canonical deliverables live in artifacts/Projects, not share URLs.
Why this matters
Share links leak context; starred chat sprawl becomes unsearchable IP graveyard.
Use stars for in-progress only; archive to Project on completion. Shared links: no client PII, rotate if leaked. Team plans: understand admin visibility policies.
Workflow — do this next
- 01Export artifact before share link.
- 02Monthly chat archive purge.
- 03Naming convention: [CLIENT]-[TASK]-[DATE].
4.7
Claude Desktop app — full setup
MCP configuration, desktop extensions (MCPB), menu bar access, and local vs remote connectors
Key takeaway
Desktop app is the MCP power surface — configure mcpServers, install Desktop Extensions from gallery, use for daily connector workflows.
Why this matters
Browser Claude.ai lacks full local MCP — Desktop is required for some enterprise integrations.
Settings → Developer → MCP servers (stdio or remote URL). Desktop Extensions: signed MCP bundles for IT distribution. Keep config in dotfiles or IT package manager.
Workflow — do this next
- 01Install Desktop; sign in work account.
- 02Add one read-only MCP server.
- 03IT packages MCPB for team if needed.
4.8
Claude Mobile — capabilities & limits
Remote MCP on mobile, voice input, photo capture, and what to defer to desktop
Key takeaway
Mobile: quick Q&A, photo capture to vision, remote MCP read — defer artifact builds, skill authoring, and local MCP to desktop.
Why this matters
Users attempt full workflows on phone — quality and security suffer.
Capture whiteboard → continue on desktop Project. Voice dictation for intake; polish on desktop. Remote connectors only — no local MCP on mobile.
Workflow — do this next
- 01Mobile for capture and triage.
- 02Sync via Project continuation.
- 03Never approve connector writes on mobile walking.
Ready-to-use artifacts
Complete templates — paste directly into your AI tool or automation workflow.
Claude.ai session startup checklist
Pin in your primary Project.
[ ] Correct Project open [ ] Web search: ON only if this thread needs live data [ ] Artifacts: ON for anything I'll export [ ] First message states: output type, artifact, source rules (if search) [ ] Sensitive data check passed
Research output QA checklist
[ ] Every number has a citation [ ] 3 random links opened and read [ ] Recency rule satisfied [ ] Conflicts with internal docs resolved [ ] Exported to wiki with date in filename
User preferences starter template
ROLE: I am a [role] at [type of org]. FORMAT: Default to [bullets / tables / short paragraphs]. Lead with the answer; put caveats after. LENGTH: [Concise / Standard / Deep dive when I ask]. Skip preamble and filler phrases. AVOID: [List 2–3 pet peeves, e.g. hedge words, unsolicited emojis].
Feature toggle matrix
TASK TYPE | SEARCH | ARTIFACTS | CODE RUN | THINKING | SKILLS Drafting / editing | OFF | optional | OFF | OFF | if needed Live market scan | ON | ON | OFF | OFF | research skill Data / spreadsheet | OFF | ON | ON | OFF | xlsx skill Hard strategy/ADR | OFF | ON | OFF | ON | optional Build mini-tool | OFF | ON (React)| ON | OFF | OFF Sensitive internal | OFF | ON | OFF | OFF | internal only
Artifact assignment brief
DELIVERABLE: [name] TYPE: [markdown | code | React | Mermaid] STRUCTURE: [sections / files] ACCEPTANCE: - [ ] Matches template - [ ] No placeholder text - [ ] Exported to [path] CHAT: Questions only — final work lives in artifact.
Search directive block
SOURCE RULES: - Prefer: [official docs, .gov, primary research] - Avoid: [affiliate blogs, unsourced forums] - Recency: [last 12 months unless historical] - If insufficient sources, say "insufficient" — do not guess
Claude.ai skills governance
□ Skill source trusted (Anthropic / partner / internal)? □ Scripts reviewed for network/exfil risk? □ Test output on synthetic data first? □ Version recorded in team registry? □ Team members know per-user upload model?
Marketing team — from chat chaos to artifact SOPs
A 9-person marketing team used Claude.ai for campaigns, research, and landing copy. Outputs lived in 200+ threads; brand voice drifted; one wrong web stat reached a press release.
Before
No shared preferences. Search always on. Copy-paste from chat. No citation review.
After
Chapter 3 rollout: team custom instructions for brand voice, artifact briefs for every deliverable, search only on 'intel' threads with SOURCE RULES block, Monday citation QA.
- Time to approved draft → down 35%
- Off-brand phrasing in reviews → down 60%
- Research errors reaching stakeholders → 2 in Q1 to 0 in Q2
- Threads used as storage → eliminated via export SOP
What goes wrong
2000-word custom instructions bloating every chat.
Move project content to Projects; keep globals under 400 words (1.2).
Treating React artifacts as production apps with PII.
Prototype only; export code to proper hosting (2.6).
Web search on for creative writing — irrelevant citations.
Toggle matrix in 1.4; default search OFF for drafting.
Sharing chat links instead of exported artifacts.
Artifact export SOP before any external share (2.7–2.8).
Vetted by Krishna KumarCurator, FactorBeam
Discussion
Discussion coming soon
Shared comments for this playbook are not live yet. When they are, you'll be able to ask questions, share what worked, and see replies from other readers.