Skip to main content
What this unit solvesThis is the primary reference for the whole Playbook. It unfolds the configuration surface of all three Anthropic products (Claude.ai web, Claude Cowork, Claude Code) and builds a complete reference for “configuration as interface.” Every layer you see here (personalization, privacy, memory, rules, permissions, automation) has a counterpart in other tools; the only differences are the names and filenames. That mapping is the work of 02-6.

​
Learning objectives

  • Configure personal instructions in Claude.ai and verify whether your conversations are used for training.
  • Explain clearly the positional difference between Claude Cowork and Claude Code without conflating the two.
  • Read and build the .claude/ directory structure, settings.json, and a multi-layer CLAUDE.md.
  • Use permissions allow / ask / deny to constrain tool permissions to task scope.
Currency noticeConfiguration details in this unit were verified as of 2026-05; sources are at the end. Anthropic renames UI elements and reorganizes file structures frequently. When in doubt, check the current official page.

​
A. Claude.ai web (Free / Pro / Max / Team)

Three settings determine whether Claude.ai is actually useful to you: personalization, privacy, and Projects. Most people only use the chat box and never visit any of these three places.

​
A1. Personalization

  • Instructions for Claude: an account-level persistent custom instruction, located at your name initials in the lower left corner → Settings. It functions as a system instruction prepended automatically to every conversation. Write your role, domain, preferred language, and output preferences once; every conversation benefits.
  • Skills: response style configuration (as of 2026-05; this feature was previously called “Styles” — refer to the current support.claude.com page). You can upload your own writing samples via “Add Writing Example” for Claude to mimic; advanced users can check “Use custom instructions (advanced)” to override directly.
  • Projects: each Project has an independent set of project instructions and a knowledge base. PDFs, DOCX files, and code uploaded to a Project are read by all conversations within that Project. Free accounts have a Project count limit (5 as of 2026-05; refer to the current support.claude.com page). For researchers: open one Project per paper topic, put the relevant literature in the knowledge base, and every conversation automatically has that context.
  • Memory: as of 2026-05, Claude maintains separate memory summaries for each Project and for non-Project conversations with cross-conversation continuity; Free plan project memory toggle details are subject to the current support.claude.com page.

​
A2. Privacy and data control

Read this section before personalization.
Set privacy before you act. This is irreversible.Before feeding research data, unpublished manuscripts, or proprietary code into Claude.ai, confirm the state of Help Improve Claude and the data terms for your plan. Once data enters training, it cannot be retroactively removed. For one-off sensitive content, use temporary conversations rather than regular ones.
Help Improve Claude toggle (Settings → Privacy): under the Free plan, this is user-controlled. Turning it off means new conversations and coding sessions will not enter future model training. Two caveats you must know: conversations flagged by safety mechanisms may still be used for trust and safety; data that has already entered a completed training run cannot be retroactively removed. Retention periods: the retention period differs between users who have opted out of training and those who have not (specific figures should be taken from the current privacy.claude.com page; this Playbook does not cite numbers that have not been officially confirmed). Team / Enterprise: governed by separate commercial terms; business data is not used for training by default. The toggle location and name differ from the Free version; refer to each plan’s admin documentation for exact names. The minimum flow for configuring this toggle:
1

Open Settings

Click your name initials in the lower left corner of Claude.ai and select Settings.
2

Navigate to the Privacy tab

Locate the “Help Improve Claude” section.
3

Confirm the toggle state matches your intent

Turn it off to opt out of training. Note: this only affects new conversations; previously submitted data is not reversible.
4

Use temporary conversations for sensitive content

For one-off sensitive documents, select the temporary conversation option next to the chat box. It does not go through the normal memory path.

​
B. Claude Cowork

Clearing up a persistent confusionAn early draft of this Playbook treated Cowork as Claude Code. They are two different products with completely different configuration models and must be understood separately.
Positioning: Anthropic’s desktop agentic product, aimed at knowledge workers (research, analysis, finance, legal, operations), completing multi-step tasks on your local files and applications. It is not a CLI and not a developer tool. Difference from Claude Code: Configuration surface: primarily a goal-oriented interface — you supply a task objective and it executes locally. There is no settings.json / CLAUDE.md configuration file system like Claude Code. Enterprise connectors (Google Drive, Gmail, DocuSign, FactSet, etc.) and domain plugins launched around early 2026 (as of 2026-05); the specific configuration mechanism for connectors is not fully indexed in publicly available official documentation; refer to anthropic.com/product/claude-cowork for the current state.

​
C. Claude Code (CLI / IDE agent, the primary reference)

This is where configuration depth is greatest, and the fullest implementation of “configuration as interface” in this Playbook. The Skill / Hook / Subagent content in 02-5 and the tool comparison in 02-6 are both extensions of this configuration model.

​
C1. The .claude/ directory structure

Cross-reference the official interactive directory explorerAnthropic provides an interactive .claude directory explorer that documents when each file is loaded, whether it should be committed, its purpose, and examples. It is the fastest way to orient yourself before getting started: Explore the .claude directory (official) (as of 2026-06). The tree below covers the commonly used subset; fields not in the tree should be taken from the official page.
Project root (your repo/, mostly committed to version control)
CLAUDE.md · project instructions, loaded every session (also works at .claude/CLAUDE.md)
CLAUDE.local.md · personal project instructions, gitignored
.mcp.json · project-level MCP servers, shared with the team
.worktreeinclude · list of gitignored files to copy into new worktrees
.claude/ · project-layer settings, rules, extensions
settings.json · project-scoped settings, committed
settings.local.json · personal overrides, gitignored
User layer (~/.claude/, Windows: %USERPROFILE%\.claude\)
CLAUDE.md · user-layer instructions, loaded every session
settings.json · applies to all projects
keybindings.json · custom keyboard shortcuts
rules/ skills/ commands/ output-styles/ agents/ workflows/ · user-layer counterpart directories
User home (~/, outside ~/.claude/)
.claude.json · app state, UI preferences, personal MCP servers, per-project trust decisions
There is exactly one criterion for deciding whether a file belongs in version control: is it a team-shared convention, or is it your personal preference or a secret? The former gets committed (settings.json, ./CLAUDE.md); the latter goes in .gitignore (settings.local.json, CLAUDE.local.md). This rule runs through all of 02-1.

​
C2. Configuration priority order

When there is a conflict, the winner from highest to lowest:
  1. Managed policy (enterprise MDM / Group Policy; cannot be overridden)
  2. CLI arguments
  3. .claude/settings.local.json (local personal)
  4. .claude/settings.json (project)
  5. ~/.claude/settings.json (user)
The closer the configuration is to the current task, the higher its priority. Managed policy is the only exception that flows from the top and cannot be violated.

​
C3. settings.json and permissions

Key fields (as of 2026-05; refer to code.claude.com/docs/en/settings): model, permissions (allow / ask / deny three arrays), hooks, env, statusLine, autoMemoryEnabled.
Permissions are the first thing to configureThe evaluation order is deny first, then ask, then allow, so deny always overrides allow. At a minimum, put secrets paths and dangerous commands in deny. This is the first line of defense described in 03-3.

​
C4. CLAUDE.md memory and imports

  • Layers from highest to lowest: managed → user (~/.claude/CLAUDE.md) → project (./CLAUDE.md) → local (./CLAUDE.local.md, gitignored).
  • Import syntax: use @path/to/file inside CLAUDE.md to import other files; paths are resolved relative to the importing file; up to 4 levels of recursion. This lets you break large rule sets into modules instead of stuffing everything into one large file (how to split them is the topic of 04-2).
  • Path-scoped rules: a .claude/rules/*.md file can include paths: ["src/**/*.ts"] in its YAML frontmatter and will only be loaded when you are working with matching files, saving context window space the rest of the time.

​
C5. Auto memory

  • Location: ~/.claude/projects/<repo-derived-path>/memory/MEMORY.md, loaded every session at up to the first 200 lines or 25 KB (whichever is hit first). Topic files (e.g., debugging.md) are not loaded at startup; they are pulled on demand when a relevant task appears. Requires Claude Code v2.1.59 or later (as of 2026-06, per the official memory documentation).
  • Control: autoMemoryEnabled in settings.json, or the environment variable CLAUDE_CODE_DISABLE_AUTO_MEMORY=1, or toggle with the /memory command.
Memory is the foothold for cross-session prompt injectionMemory is loaded at session start. Keep its contents narrow and never store secrets in it (see 03-3).

​
C6. Windows specifics

  • User path: C:\Users\<user>\.claude\ (that is, %USERPROFILE%\.claude\).
  • Managed policy (as of 2026-05, refer to code.claude.com/docs): C:\Program Files\ClaudeCode\managed-settings.json. The exact registry key paths used for Group Policy / Intune deployment are documented in the enterprise deployment section of code.claude.com/docs/en/settings.

​
Tool comparison: three Anthropic products at a glance

This unit focuses on Claude. The table below is a positioning summary of Anthropic’s own three products, not a cross-vendor concept mapping. The standard cross-vendor table (OpenAI / Google / Copilot / Cursor, concept row by row) is in 02-6:

​
Hands-on exercises

1

Configure Claude.ai personalization and privacy

Set Instructions for Claude (role, language, output preferences) and go to Settings → Privacy to confirm the state of Help Improve Claude.
2

Create a minimal working CLAUDE.md

In a project, create ./CLAUDE.md with five sections: role, language, code style, constraints, and explicit prohibitions (template in 04-1).
3

Verify a deny rule

Add one deny rule to .claude/settings.json (e.g., Read(./.env)) and deliberately trigger it to confirm it blocks.
4

Move personal preferences to the user layer

Move a cross-project personal preference to ~/.claude/CLAUDE.md and confirm it takes effect in a different project.

​
Common pitfalls

Anti-patterns
  • Treating Cowork as Code or vice versa, leading to a fruitless search for configuration that does not exist there.
  • Writing personal preferences into the team-shared ./CLAUDE.md, contaminating colleagues’ environments. Personal preferences belong in ~/.claude/CLAUDE.md or CLAUDE.local.md.
  • Forgetting to add settings.local.json to .gitignore, committing personal overrides alongside shared configuration.
  • Feeding sensitive data before confirming privacy settings, then discovering the action is irreversible.

​
Self-check

The bar for passing this unit
  1. Can you draw out the .claude/ directory and explain for each file whether it should be committed and why?
  2. Do you know whether your Claude.ai account is currently contributing conversations to training? Where is the toggle?
  3. Given a path “you do not want Claude to touch,” can you write the corresponding deny rule?

​
Sources and further reading

Factual claims are grounded in official documentation; fast-changing items are annotated as of 2026-05.
  • Related: 02-1 configuration layer model and priority order; 02-5 deep dive into Skill / Hook / Subagent / Plugin; 02-6 cross-tool concept mapping (OpenAI / Google / Copilot / Cursor); 03-3 security, privacy, and supply chain threat model; 04-1 CLAUDE.md memory design; 04-2 modular rule splitting.
  • Fast-changing items (retention period figures, Team / Enterprise toggle names, Cowork connector configuration) have been annotated with their respective official pages; re-verify against the current pages before this becomes formal course material.