respond

antorsae's avatarfrom antorsae

Manually write response back to Claude. Use if read-task did not auto-complete or to finalize response.

⭐2stars🔀0forks📁View on GitHub🕐Updated Jan 3, 2026
[Tool Routing and Orchestration]

When & Why to Use This Skill

The Respond Skill is a critical utility for Claude agentic workflows, designed to facilitate manual communication between an agent and the orchestrator. It ensures that task findings, code implementations, and analysis are properly documented and signaled as complete, especially in scenarios where automated triggers fail. By providing a structured markdown template for responses, it maintains consistency and reliability in multi-step agent collaborations.

Use Cases

  • Case 1: Manual Task Finalization - Use this skill to explicitly send findings back to Claude when a background task or file operation does not automatically trigger a completion signal.
  • Case 2: Structured Technical Reporting - Ideal for formatting complex outputs like code reviews (CODE_REVIEW) or implementation plans (IMPLEMENT) into a standardized markdown structure for better readability and downstream processing.
  • Case 3: Workflow Recovery - Acts as a fail-safe mechanism to update the agent's status to 'done' and provide a summary of work when the standard execution flow is interrupted or requires human-in-the-loop intervention.
namerespond
descriptionManually write response back to Claude. Use if read-task did not auto-complete or to finalize response.

Respond Skill

Manually write response back to Claude.

When to Use

  • If /read-task didn't auto-complete
  • To finalize or modify a response
  • To manually send findings to Claude

Steps

Before any file operations, resolve the .agent-collab directory so commands work outside the project root:

AGENT_COLLAB_DIR="${AGENT_COLLAB_DIR:-}"
if [ -n "$AGENT_COLLAB_DIR" ]; then
  if [ -d "$AGENT_COLLAB_DIR/.agent-collab" ]; then
    AGENT_COLLAB_DIR="$AGENT_COLLAB_DIR/.agent-collab"
  elif [ ! -d "$AGENT_COLLAB_DIR" ]; then
    AGENT_COLLAB_DIR=""
  fi
fi

if [ -z "$AGENT_COLLAB_DIR" ]; then
  AGENT_COLLAB_DIR="$(pwd)"
  while [ "$AGENT_COLLAB_DIR" != "/" ] && [ ! -d "$AGENT_COLLAB_DIR/.agent-collab" ]; do
    AGENT_COLLAB_DIR="$(dirname "$AGENT_COLLAB_DIR")"
  done
  AGENT_COLLAB_DIR="$AGENT_COLLAB_DIR/.agent-collab"
fi

If $AGENT_COLLAB_DIR does not exist, stop and ask for the project root.

1. Gather Response

If task not yet completed, finish it now.

Collect all findings, code, or analysis.

2. Write Response

Write to $AGENT_COLLAB_DIR/responses/response.md:

# Codex Response

## Task Type
[CODE_REVIEW | IMPLEMENT | PLAN_REVIEW]

## Completed At
[Current timestamp]

## Summary
[Brief summary]

## Detailed Findings/Output
[Full response content]

3. Update Status

Write done to $AGENT_COLLAB_DIR/status

4. Confirm

Tell user: "Response written. Claude can now use /codex-read to retrieve it."