This lesson is being recorded

Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.

Bruno AshfordBruno Ashford @brunoashfordSubscribe

Module 11 · Lesson 11.1 · 11 min

Permission Modes

Understand Claude Code's permission modes and rules, and configure how much the agent can do without asking you.

Module 11
Security, Governance and Trust

Real power for your agents, without losing sleep

Bruno Ashford
Bruno Ashford · Module 1101 / 12
The principle
Trust is not a feeling. It is boundaries you set, checks you run and evidence you keep.
Bruno Ashford
Bruno Ashford · Module 1102 / 12
Lesson 11.1

Permission modes

ModeWhat runs without askingUse it for
Manual (default)Reads onlySensitive work
acceptEditsReads and file editsDrafting you review
planReads, no changesThinking before acting
autoMost actions, classifier reviewsLong tasks
dontAskOnly pre-approved toolsScheduled runs
bypassPermissionsEverythingIsolated containers only

Press Shift+Tab to cycle. Check the status bar.

Bruno Ashford · Module 1103 / 12
Lesson 11.1

Rules beat reminders

Deny beats ask beats allow, in every mode.

  • Allow the boring, safe work
  • Ask for anything sensitive
  • Deny secrets, deletes and publishing
  • Review them with /permissions
Bruno Ashford · Module 1104 / 12
Lesson 11.2

Customer data habits

  1. 1List itKnow what personal data you share
  2. 2Minimise itOnly the columns the task needs
  3. 3Separate itOne client, one folder
  4. 4Check termsKnow your plan's data settings
Bruno Ashford · Module 1105 / 12
Lesson 11.2

The laws in the background

Good habits, not legal advice.

  • UK GDPR and EU GDPR
  • US state privacy laws, like California's
  • Sector rules for health, legal, finance
  • Speak to a lawyer for specifics
Bruno Ashford · Module 1106 / 12
Lesson 11.3
A secret that has been in a chat or a synced file should be treated as exposed.
Bruno Ashford
Bruno Ashford · Module 1107 / 12
Lesson 11.4

Who approves what

ActionLevelApprover
Draft a replyAutomaticNone
Send to a clientApprovalAccount owner
Refund over limitApprovalOwner
Delete recordsNever delegatedOwner, by hand
Legal noticeNever delegatedOwner and lawyer

Your full matrix is in the materials.

Bruno Ashford · Module 1108 / 12
Lesson 11.5

Three layers of verification

  1. 1Self-checkAgainst a short checklist
  2. 2Separate reviewerFresh session or subagent
  3. 3Human sample100 percent first, then sample
Bruno Ashford · Module 1109 / 12
Illustrative example, not research

One law practice, one month apart

9 to 1errors found per month
100% to 25%letters reviewed by a person
0wrong letters sent to clients
Bruno Ashford · Module 1110 / 12
Lesson 11.6

A policy people actually read

One page, plain English.

  • Approved tools and accounts
  • What data can and cannot be used
  • What always needs approval
  • Who to ask, and review quarterly
Bruno Ashford · Module 1111 / 12
Recap
Governance is what lets you say yes to AI with confidence, instead of hoping nothing goes wrong.
Bruno Ashford
Bruno Ashford · Module 1112 / 12
01 / 12

Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.

1Know the modes

Manual mode (config value default) asks before edits and commands. acceptEdits lets Claude edit files without asking. plan mode lets it read and plan but not change anything. auto mode uses a second model, a classifier, to review actions in the background instead of you. dontAsk denies anything not pre-approved, ideal for scheduled runs. bypassPermissions skips checks entirely and belongs only in isolated containers.

2Switch modes during a session

Press Shift+Tab to cycle through the modes; the status bar at the bottom shows which one is active. In recent versions many sessions start in auto mode, so check the status bar rather than assuming. Use plan mode whenever you want Claude to think before touching anything, such as a new process or a sensitive client file.

Terminal
claude --permission-mode plan
3Pre-approve the safe and ban the dangerous

Permission rules live in .claude/settings.json under allow, ask and deny. Deny always wins, then ask, then allow, and deny rules apply in every mode, even when checks are otherwise skipped. Type /permissions to see and edit your rules and which file each one comes from.

Terminal
{
  "permissions": {
    "allow": [
      "Read",
      "Edit(./reports/**)",
      "Bash(git status *)"
    ],
    "ask": [
      "Bash(git commit *)"
    ],
    "deny": [
      "Read(./.env)",
      "Read(./secrets/**)",
      "Bash(git push *)",
      "Bash(rm -rf *)"
    ]
  }
}
4Set a default for your project

You can make a mode the starting point for a project by setting defaultMode inside permissions, for example plan for a folder with sensitive client files. Personal preferences go in ~/.claude/settings.json; team rules go in the project's .claude/settings.json so everyone gets them.

Prompt for Claude
Review my .claude/settings.json. Propose permission rules for this business folder: allow reading everything except secrets, allow edits only inside reports and drafts, ask before any git commit, and deny reading .env files, deleting folders and pushing to any remote. Explain each rule in one plain-English line before changing anything.
5Remember what rules can and cannot do

Permission rules are enforced by Claude Code itself, not by the model, which is exactly why they are reliable. Instructions in CLAUDE.md shape what Claude tries to do, but they do not change what it is allowed to do. For anything that must never happen, use a deny rule or a hook, not a sentence in a prompt.

Owners tend to fall into one of two traps. Some approve every single action for weeks, get tired, and start clicking yes without reading, which is worse than no checks at all. Others switch everything off on day two because the prompts slowed them down. Both are the same mistake: treating permissions as all or nothing.

Claude Code gives you a dial, not a switch. Permission modes set the baseline for a session, and permission rules let you pre-approve safe actions and permanently ban dangerous ones. The goal is simple: no prompts for the boring, safe work, and a hard stop for anything that can hurt the business.

Key points

  • Permissions are a dial, not a switch.
  • Shift+Tab cycles modes; the status bar tells you which is active.
  • Deny beats ask beats allow, in every mode.
  • Rules are enforced by Claude Code; prompts are only guidance.

Avoid

  • Approving everything without reading because the prompts feel repetitive.
  • Using bypassPermissions on your own laptop.
  • Writing never delete files in CLAUDE.md instead of a deny rule.
  • Putting team rules in personal settings, so nobody else gets them.
Real-world example: an accounting practice

A three-partner accounting firm wanted staff to use Claude Code on client folders but worried about tax records being altered. They set plan mode as the default for the clients folder, allowed edits only in a drafts folder, and denied reading any file with credentials in its name. Staff now draft letters and reconciliations freely in drafts, and nothing in the original client records can be changed without a person moving it there. Approval prompts dropped by roughly 70 percent while the partners slept better.

Bruno Ashford Subscribe