Stay In Lane

SkillFiles & storage

Before making changes, verifies they match what was actually requested. Activates when about to modify files, add features, or refactor code. Catches scope creep before it happens - no "while I'm here" improvements.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Stay In Lane skill

What this skill tells your AI

The instructions your AI receives, as published by elliotjlt/claude-skill-potions in skills/stay-in-lane/SKILL.md and read by ahel’s review.

When To Activate

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

MatchAction
YesProceed
NoStop. Don't do it.
AdjacentAsk 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)

  • "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.

Signals

GitHub stars
62
Forks
2
Last commit
Feb 2026
Advanced
Catalog kind
skill
Gateway key
stay-in-lane
Source
github.com/elliotjlt/claude-skill-potions