Developer Experience

Amplifier Modes

Runtime behavior overlays that modify how the assistant operates — through context injection, tool moderation, and visual feedback

Active

February 2026

The Problem

Planning and implementing don't mix

Sometimes you want to think through a problem before writing code. But the assistant doesn't know that.

Premature Implementation

You ask about architecture and get 500 lines of code. You wanted a discussion, not a commit.

Manual Reminders

"Don't write any files yet" repeated in every message. Tedious and easy to forget.

No Feedback

No visual indication of current operating mode. Is the assistant planning or implementing? No way to tell.

The Solution

Behavioral overlays

Not just "block tools" — guide the assistant's mindset

Context

System Reminder

+

Enforcement
Tool Moderation

+

Visibility

Prompt Indicator

            Modes combine context injection, tool policies, and visual feedback for consistent behavior control. Mechanism, not policy.
        

Built-in Modes

Three modes for different workflows

/plan

Analyze, strategize, and organize — but don't implement. Read-only tools plus agent delegation. File writes blocked entirely.

/explore

Zero-footprint codebase exploration. Only read_file, glob, grep, web search, and LSP allowed. Everything else blocked.

/careful

Full capability with confirmation gates. Write operations and bash require explicit user approval before execution.

Enforcement

Four-tier tool policy system

Safe

read_file, glob, grep, LSP, web_search

Always allowed. Reading and analysis tools work normally.

Warn

bash (in plan mode)

First call blocked with warning. Retry to proceed after justification.

Confirm

write_file, edit_file, bash (in careful mode)

Requires explicit user approval via the approval hook before execution.

Block

Any unlisted tool (default_action)

Denied entirely. Tool call fails with mode policy explanation.

Visual Feedback

Always know your mode

The prompt changes to show the active mode

[plan]> describe the authentication flow...

Without Modes

> "Remember, don't write files yet, just discuss the architecture..."

Repeated every message. Easy to forget.

With Modes

[plan]> "How should we structure authentication?"

Mode is visible. Behavior is enforced.

Architecture

Five-step runtime pipeline

1

Discovery

ModeDiscovery searches 5 paths: project, user, bundle, config, composed bundles

2

Loading

parse_mode_file() extracts YAML frontmatter config + markdown context text

3

Context Injection

provider:request hook injects mode guidance as <system-reminder source="mode-*">

4

Tool Moderation

tool:pre hook checks each tool call against the mode's safe/warn/confirm/block policy

5

Approval Integration

"confirm" tools delegated to hooks-approval via require_approval_tools in session state

Interface

Simple commands

/modes

List all available modes with descriptions

/mode <name>

Toggle a mode on/off (e.g., /mode plan)

/mode off

Clear any active mode and return to normal

/plan /explore /careful

Shortcuts — activate built-in modes directly

            Agent-initiated too: The mode tool exposes set, clear, list, current operations. Default gate policy is "warn" to prevent accidental transitions.
        

Ecosystem

13 modes across 3 bundles

Any bundle can contribute modes by placing .md files in a modes/ directory

3

amplifier-bundle-modes

plan, explore, careful

6

amplifier-bundle-superpowers

brainstorm, write-plan, execute-plan, verify, finish, debug

4

amplifier-bundle-daily-flow

brief, dispatch, eod, idea

Extensible

Create custom modes via markdown

amplifier-bundle-modes/ ├── bundle.md # Thin wrapper ├── behaviors/modes.yaml # Hook config ├── modes/ # Built-in modes │ ├── plan.md # 58 lines │ ├── explore.md # 53 lines │ └── careful.md # 48 lines └── modules/ ├── hooks-mode/ # 613 lines Python └── tool-mode/ # 319 lines Python

                    Create your own: Add a markdown file to .amplifier/modes/ or ~/.amplifier/modes/. Define YAML frontmatter with name, description, tool policies, and default_action. The markdown body becomes context injection text.
                

                    Philosophy: Mechanism not policy. The bundle provides generic hooks, discovery, and moderation. Mode files provide the policy. Users customize freely without writing code.
                

Development

Implementation snapshot

2,622

Total lines (source files)

2,045

Lines of Python

2

Hook modules

3

Built-in modes

            Repository: microsoft/amplifier-bundle-modes

            Primary contributor: Brian Krabach (sole author in visible git history)

Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Active

Repository: microsoft/amplifier-bundle-modes

Research performed:

Bundle discovery: find ~/.amplifier/cache -path "*modes*" -type f (20+ files found)
Mode definitions read: cat modes/plan.md, modes/explore.md, modes/careful.md
Bundle config: cat behaviors/modes.yaml (hooks-mode + hooks-approval + tool-mode)
Git log: git log --oneline (1 commit visible in shallow clone)
Contributors: git log --format="%an" | sort | uniq -c → Brian Krabach (1 commit)
Line counts: find . -type f -name "*.py" | xargs wc -l → 2,045 lines Python
Source files: find . -not -path './.git/*' -type f | xargs wc -l → 2,622 lines total
Ecosystem scan: find ~/.amplifier/cache -name "*.md" -path "*/modes/*" → 13 mode files across 3 bundles
Tool-mode source: head -80 modules/tool-mode/amplifier_module_tool_mode/__init__.py (confirmed set/clear/list/current operations)
Remote: git remote -v → https://github.com/microsoft/amplifier-bundle-modes

Gaps: Shallow clone limits git history to 1 commit. Full commit count, development timeline, and complete contributor list unavailable. Bundle.md references a "review" mode in examples but no review.md exists — the actual third mode is "careful".

Primary contributor: Brian Krabach (100% of visible commits)

Try It

One command changes everything

/

/plan — Think before implementing

/

/explore — Understand before acting

/

/careful — Confirm before changing

[plan]> let's think this through...