Amplifier Stories

Recipe Runner Mode

Active

A Story of Human-AI Collaboration

January 2026

1

Act 1

The Problem

When agents improvise instead of following instructions

The Scenario

Authoring a PR Review Recipe

A user was writing a recipe to automate code reviews...

# pr-review.yaml (with typos and errors)
name: "PR Review Recipe"
steps:
  - action: "fetch_pr_detials"  # typo!
    with:
      pr_number: ${context.pr}  # wrong syntax
  - action: "analyz_code"  # another typo

What Happened

Agent Improvisation

Expected

"Your recipe has errors on lines 4 and 7. Here's how to fix them..."

Report the problem, let user fix it

Actual

"I'll fetch the PR details for you and analyze the code..."

Agent read the recipe and executed the steps itself

The agent tried to be helpful by doing what the recipe described, instead of running the recipe through the proper system.

The Frustration

"I want the recipe to run, not for the agent to improvise and do it itself."

The CLI tool existed, but users wanted a human way to say:
"Run this recipe, but that's all - don't do it yourself"

2

Act 2

The Solution Emerges

Amplifier suggests the "mode" approach

The Conversation

Amplifier as Architect

User describes the need

"I want to restrict what the agent can do when I'm testing recipes..."

Amplifier suggests modes

"Modes are runtime behavior overlays that restrict what tools an agent can use. You could create a mode that only allows the recipes tool."

Collaborative design

Together, they designed the recipe-runner mode specification

The Solution

.amplifier/modes/recipe-runner.md

# Recipe Runner Mode

tool_policies:
  default_action: block   # Block everything by default

tools:
  recipes: safe            # Only allow the recipes tool

Default: Block

All tools are blocked unless explicitly allowed. No file edits, no bash, no web access.

Exception: Recipes

The recipes tool is marked as safe - the agent can execute, resume, and manage recipes.

The Difference

Before vs. After

Without Mode

# User: "Run my PR review recipe"

Agent thinks: "I'll help with PR review!"
Agent does:
  - Reads the recipe file
  - Tries to fetch PR details itself
  - Calls git, reads files, etc.
  - Improvises the workflow

With Recipe-Runner Mode

# User: "Run my PR review recipe"

Agent thinks: "I can only use recipes tool"
Agent does:
  - Calls recipes(execute, "pr-review.yaml")
  - Recipe system validates YAML
  - Reports errors if any
  - Proper recipe execution

3

Act 3

Demo Day

Sharing the solution with the team

Thursday Meeting

Team Demo: Recipe Runner Mode

+1

Great Reception

Team saw immediate value in restricted execution modes

?

A Question Arose

"Can I share this mode with others?"

"Can I share this mode in a bundle?"

Architectural Guidance

Where Should It Live?

User asks

"Can I share this mode in a bundle? Maybe the PR review bundle?"

Amplifier advises

"I'd suggest the recipes bundle instead - this mode is general-purpose, not specific to PR reviews."

User agrees

"Good call - better architecture."

Amplifier as architectural advisor - not just executing commands, but providing guidance on system design and organization.

4

Act 4

The Plot Twist

A bug lurks in the mode system

Bug Discovery

When Adding to the Bundle...

# Trying to use the bundled mode
$ amplifier --mode @recipes:recipe-runner

Error: Tool 'recipes' is blocked by current mode policy

# Wait, what? The mode explicitly allows recipes!

Problem 1

Mode system couldn't discover modes from other bundles

Problem 2

The recipes tool was being blocked despite being marked safe

Root Cause

Schema Mismatch

Mode Config Provided

tool_policies:
  default_action: block
tools:
  recipes: safe

Code Expected

tool_policies:
  default_action: block
  tools:           # nested!
    recipes: safe

The hooks-mode code expected tools to be nested under tool_policies,
but the mode file had it at the top level.

5

Act 5

The Fix

During a meeting, with minimal attention

Minimal Attention Required

User Was in a Team Meeting

The user described the bug briefly and let Amplifier work autonomously:

Investigate

Found the session where the bug occurred

Diagnose

Identified the schema mismatch in hooks-mode

Fix

Updated parser to handle both schemas

Enhance

Implemented cross-bundle mode discovery

Test

Used shadow environments to verify

3

User prompts needed

Complex debugging and fix
with minimal human attention

6

Act 6

The Shadow
Environment Hero

Safe testing for dangerous changes

Safe Testing

Why Shadow Environments Mattered

The fix involved changes to core mode/hook code - the very system that controls tool access. Testing this in production was risky.

Isolation

Changes tested in an isolated container, separate from the main environment

Verification

Confirmed the fix worked before committing to the real codebase

Regression Check

Caught potential issues that could break other modes

Code Change

→

Shadow Env

→

Test Mode

→

Commit

Takeaways

Key Themes

🤝

Human-AI Collaboration

Amplifier as architectural advisor, not just code executor

💡

Minimal Attention

Complex debugging with just a few user prompts

📦

Extensibility

Modes can be shared in bundles - community benefit

🛡

Safety

Shadow environments for testing changes to core systems

The Journey

Organic Problem Solving

Problem

→

Discussion

→

Solution

→

Sharing

→

Bug

→

Fix

This wasn't a planned feature. It emerged from a real user frustration, evolved through collaboration, and improved the system for everyone.

Velocity

By the Numbers

1

Mode created

2

Bugs fixed

3

User prompts

1

Meeting multitasked

From frustration to shared solution in a single conversation,
with autonomous bug fixing during a meeting.

Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Active

Research performed:

Modes in superpowers bundle: found write-plan.md, execute-plan.md, finish.md, brainstorm.md, verify.md, debug.md in ~/.amplifier/cache
No recipe-runner specific mode files found in cache search
Narrative based on user-reported experience with recipe authoring and mode system

Gaps: Exact commit dates, PR numbers, and line counts not available from local research. Story is based on qualitative user experience narrative.

Primary contributors: samschillace (mode design and bug fixing)

Try It

Use Recipe Runner Mode

# Use the bundled mode
$ amplifier --mode @recipes:recipe-runner

# Or create your own restricted modes
$ cat .amplifier/modes/my-mode.md

Available in the recipes bundle