Claude Code Plugins

Community-maintained marketplace

Feedback

Emergency Stop (STOP Keyword)

@cipherstash/cipherpowers
0
0

Immediate halt protocol when user detects problems requiring urgent intervention

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name Emergency Stop (STOP Keyword)
description Immediate halt protocol when user detects problems requiring urgent intervention
when_to_use when user sends message starting with "STOP" keyword (all caps, first word, followed by newline)
version 1.0.0

Emergency Stop Protocol

Overview

Users can signal emergency stop using the STOP keyword to immediately halt current work when they detect problems requiring urgent intervention.

Core principle: Distinguish between queued work (finish current task first) and emergency stops (halt immediately to prevent wasted effort or propagating errors).

Quick Reference

STOP keyword pattern:

STOP[punctuation optional]
[rest of message explaining the issue]

Examples that trigger emergency stop:

STOP
You're modifying the wrong file

STOP!
That approach won't work with our architecture

STOP:
Let me explain the requirement differently

Examples that DO NOT trigger:

  • "We should STOP supporting IE11" (not first word)
  • "Please STOP this" (not first word)
  • "stop working on this" (lowercase)
  • "Let's STOP and reconsider" (not first word)

Implementation

Detection Algorithm

When receiving user message, run this algorithm:

Step 1: Check: Is first word of message "STOP" (all caps)?
        → YES: Go to Step 2
        → NO: Go to Step 8 (normal message handling)

Step 2: Check: Does "STOP" have optional punctuation (STOP, STOP!, STOP:, etc)?
        → YES or NO: Go to Step 3 (punctuation allowed)

Step 3: Check: Is there content after the first line?
        → YES: Go to Step 4 (emergency stop)
        → NO: Go to Step 4 (emergency stop - user will explain)

Step 4: IMMEDIATE HALT
        - Stop current work immediately (mid-task is fine)
        - Do NOT complete current function/test/file/batch
        - Save mental state of what you were doing
        Go to Step 5

Step 5: Acknowledge emergency stop
        Output: "[EMERGENCY STOP] Work halted. What's the issue?"
        Go to Step 6

Step 6: Read and understand user's concern
        - User message after "STOP" explains the problem
        - Do NOT assume what's wrong
        - Do NOT offer solutions yet
        - Focus on understanding the issue
        Go to Step 7

Step 7: Wait for user direction
        - Ask clarifying questions if needed
        - Wait for user to provide new direction
        - Do NOT resume previous work unless user explicitly says to
        STOP

Step 8: Normal message handling
        - Evaluate if message is queued work vs normal response
        - If adding new task: Acknowledge and queue it
        - If question/comment: Respond normally
        Continue with current work if applicable
        STOP

Why First Word + All Caps + Newline?

Eliminates false positives:

  • "We should STOP supporting X" → NOT first word, doesn't trigger
  • "Please STOP" → NOT first word, doesn't trigger
  • "stop this" → Lowercase, doesn't trigger
  • Mid-sentence STOP → Not first word, doesn't trigger

Clear signal:

  • User must deliberately start message with STOP
  • All caps = visual emphasis
  • First word = unambiguous pattern matching
  • Newline = explanation follows

Pattern is memorable:

STOP
[explain the problem]

When User Should Use STOP

Emergency situations:

  • Agent is working on wrong feature/file/approach
  • Agent misunderstood requirements
  • User noticed a critical flaw in current direction
  • Need to provide essential context before continuing
  • About to waste significant effort if not stopped

NOT for:

  • Adding new tasks to queue ("also please do X later")
  • General questions during work
  • Normal feedback or suggestions
  • Requesting status updates

Mental model: STOP = pulling the Andon cord in Toyota factory. Only pull when defect detected, not for routine communication.

Integration with Other Workflows

During Plan Execution

If user sends STOP during /execute batch execution:

  1. Halt current task immediately (don't finish it)
  2. Acknowledge: "[EMERGENCY STOP] Batch execution halted at Task N"
  3. Listen to user's concern
  4. Ask: "Should we abandon this plan, or resume after addressing the issue?"

During Code Review

If user sends STOP during code review:

  1. Halt review immediately
  2. Acknowledge: "[EMERGENCY STOP] Code review halted"
  3. User may be pointing out critical issue missed
  4. Listen and respond to concern

During TDD Implementation

If user sends STOP during test-first implementation:

  1. Halt immediately (even mid-test)
  2. Acknowledge: "[EMERGENCY STOP] TDD cycle halted"
  3. User may have noticed wrong test case or approach
  4. Listen to concern before resuming

What NOT to Do

NEVER:

  • Finish current function "real quick" before stopping
  • Complete current test before stopping
  • Finish current batch before stopping
  • Rationalize: "I'm almost done with this part"

Rationalization defense:

  • "Let me just finish this function" → NO, stop NOW
  • "I'm 90% done with this test" → NO, stop NOW
  • "One more file and I'll stop" → NO, stop NOW
  • "I'll stop at the next checkpoint" → NO, stop NOW

Why: If user says STOP, continuing wastes effort. The work might be wrong, the file might be wrong, the approach might be wrong. Better to halt immediately than compound the error.

Testing This Skill

See test-scenarios.md for pressure tests validating this protocol.

Test scenarios:

  1. Agent mid-batch, user sends "STOP\nWrong file" → Does agent halt immediately?
  2. Agent writing test, user sends "We should STOP using X" → Does agent continue (not first word)?
  3. Agent in flow, user sends "STOP!\nI need to explain something" → Does agent stop and listen?
  4. User sends "Please STOP this approach" → Does agent continue (not first word)?

Related Skills

Why this skill exists:

  • Complements "Responding to User Interrupts" (Insight 2)
  • Provides clear signal vs ambiguous "please review this"
  • Prevents "momentum mode" from ignoring urgent input
  • Based on "Jidoka" insight (Insight 3)

See also:

  • executing-plans/SKILL.md - Plan execution workflow
  • systematic-debugging/SKILL.md - When to stop and reassess

Origin

Lineage: This skill emerged from Insight 3 in docs/skill-insights-scratchpad.md.

Problem identified: Agents struggle to distinguish "queue this work" from "emergency stop."

Solution: Unambiguous keyword pattern (first word, all caps, newline) that user can use to signal urgent halt.

Etymology: "Jidoka" (自働化) - Toyota Production System principle of stopping the production line when defect detected. We use "STOP" for clarity, but the concept is Jidoka.