Creep Guard

Creep Guard

Keeps Claude focused on the stated task. When something outside the original request is detected, pause and ask — don't silently do it.

Modes

• proactive (default): Monitor autonomously throughout any task. Interrupt

whenever drift is detected.

• reactive: User invokes /creep-guard before a task. Hold that task

statement as the source of truth and check against it as you work.

Setting the Mode

• Persistent default: edit the mode: field in this file's frontmatter

(proactive or reactive). This applies whenever the skill auto-triggers or is invoked with no argument.

• Per-task override: invoke with the mode name as the argument, e.g.

/creep-guard reactive or /creep-guard proactive. This overrides the frontmatter default for the current task only and does not change the file.

• Precedence: if an argument is given at invocation, use it. Otherwise,

fall back to the frontmatter mode: value.

Source of Truth

The scope is defined by the user's prompt for the current task only. No other documents, prior conversations, or inferred intent. If the task wasn't stated explicitly, ask for clarification before starting.

When to Flag

Flag something as potential scope creep only when both are true:

1. It was not mentioned in the original task prompt
2. It would require touching files or logic outside what the task already requires

Do not flag:

• Fixing a syntax error in a line you're already editing
• Adding a missing import required by code you were asked to write
• Handling an obvious error case in a function you were asked to create

Do flag:

• Refactoring something adjacent that "looks messy"
• Adding auth, validation, or error handling to a system you weren't asked to touch
• Creating new files, modules, or abstractions beyond what the task requires
• Upgrading dependencies while in the neighborhood
• Fixing bugs you noticed that are unrelated to the task
• Adding tasks to a plan or .md file that was beyond what was requested by the user

The Interaction

When scope creep is detected, stop and say:

> "I noticed I could also [specific thing]. This wasn't part of the original task. > What would you like to do? > > - Now — do it as part of this task > - Later — log it to tech_debt.md and skip it for now > - Never — ignore it and move on"

Wait for the user's response before continuing. Do not proceed with the out-of-scope change while waiting.

On "Later": Writing to tech_debt.md

Append to tech_debt.md in the project root (create it if it doesn't exist):

- [ ] [YYYY-MM-DD] [one-line description of what was deferred] — deferred during: "[brief summary of original task]"

Keep entries short and actionable. The file is append-only — never rewrite or reorganize existing entries.

Confidence

Only interrupt if you are reasonably confident something qualifies. A low-confidence hunch is not worth breaking the user's flow. When genuinely uncertain, lean toward flagging — one unnecessary check is less costly than silently expanding scope.

Source & license

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

• Author: robbytables
• Source: robbytables/claude_skills
• License: Unlicense

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