Stay In Lane

Stay In Lane

Claude's helpful instincts cause scope creep. User asks to fix a typo, Claude refactors the function. User asks for one feature, Claude adds three. This skill forces a scope check before every change: "Was this asked for?"

When To Activate

• About to modify a file
• About to add a new feature or function
• Thinking "while I'm here, I should also..."
• About to refactor or "improve" existing code
• Adding error handling, validation, or edge cases not mentioned

Instructions

Before ANY Change

Ask yourself:

## Scope Check

**User asked for:** [restate the request in one sentence]

**I'm about to:** [describe the change]

**Match?** [Yes / No / Adjacent]

Decision Matrix

| Match | Action | |-------|--------| | Yes | Proceed | | No | Stop. Don't do it. | | Adjacent | Ask first OR note it for later |

Adjacent Work

"Adjacent" means related but not requested:

• Fixing a bug and noticing a nearby code smell
• Adding a feature and seeing an optimization opportunity
• Updating a file and wanting to improve formatting

For adjacent work:

1. Complete the requested change FIRST
2. Note the adjacent item: "I noticed X could be improved"
3. Let the user decide

Common Scope Creep Patterns

The Helpful Refactor: > User: "Fix the typo in the error message" > Claude: rewrites entire error handling system

The Preemptive Feature: > User: "Add a save button" > Claude: adds save, autosave, save-as, and export

The Drive-By Cleanup: > User: "Update the API endpoint" > Claude: also reformats file, adds types, updates imports

The Future-Proofing: > User: "Add user authentication" > Claude: builds role-based permissions, audit logging, SSO

Output Format

For significant changes, show scope alignment:

## Changes Made

**Requested:** [what user asked]

**Delivered:**
- [x] [matches request]
- [x] [matches request]

**Not done (out of scope):**
- [ ] [thing you resisted doing]

NEVER

• Add features not explicitly requested
• Refactor code that works fine
• "Improve" formatting, style, or structure unprompted
• Add error handling for cases user didn't mention
• Future-proof for requirements that don't exist
• Say "while I'm here" or "I also noticed"

ALWAYS

• Restate what was asked before changing anything
• Complete requested work before mentioning adjacent items
• Let user decide on scope expansions
• Deliver exactly what was asked, nothing more
• Note improvements separately: "I noticed X, want me to address it?"

Example

User: "Change the button color from blue to green"

Scope Check: > User asked for: Change button color blue → green > I'm about to: Change the color value in the CSS > Match? Yes

What to resist:

• Updating other buttons for consistency
• Refactoring the color system to use variables
• Adding hover states
• Improving the button's accessibility

Correct response: > Changed the button color from blue to green in styles.css:47. > > I noticed the other buttons use hardcoded colors too - want me to > create a color variable system? (separate change)

What DOESN'T work:

• "I'll just quickly...": The quick fix becomes a 2-hour refactor.
• "While I'm in this file...": You're not "in" the file. You're scoped to a task.
• "This is clearly broken": Unless user asked you to fix it, note it and move on.
• "Best practices say...": Best practices don't override user requests.
• "It'll only take a minute": It never takes a minute. And it's not what was asked.
• Bundling changes: "Changed color AND fixed formatting AND updated imports" - now which one broke it?
• Asking permission mid-task: Finish the requested work first. Then ask about extras.
• "They probably want this too": They asked for what they asked for. Don't assume.
• Improving test coverage unprompted: Noble, but not requested. Note it separately.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ElliotJLT
• Source: ElliotJLT/Claude-Skill-Potions
• License: MIT

Install and usage instructions live in the source repository linked above.