handoff

name

handoff

description

Generate optimized handoff prompts for delegating tasks to LLM agents. Use when needing to hand off work to another agent (GPT-5.2/Codex, Claude Opus 4.5/Sonnet 4, or Gemini 3 Pro/Flash), whether for a sub-task/parallel task within the same project or a fresh start handoff to a new agent context. Triggers on requests like "create a handoff prompt", "delegate this task to another agent", "hand this off", or "prepare context for another agent".

Handoff Prompt Generator

Generate optimized handoff prompts for different LLM agents and handoff types.

Handoff Types

1. Sub-Task/Parallel Handoff

For delegating a portion of work to another agent within the same project:

• Agent has access to same codebase/files
• Shared context exists
• Task is scoped subset of larger work
• May run in parallel with other agents

2. Fresh Start Handoff

For handing off to an agent starting with a new context:

• Agent starts with no prior context
• Needs project orientation
• Requires state files and entry points
• May be a new session or different model

Workflow

1. Identify target model - Ask which model family will receive the handoff
2. Identify handoff type - Sub-task/parallel or fresh start
3. Gather context - Collect essential information for the handoff
4. Generate prompt - Apply model-specific patterns from references
5. Review and refine - Ensure prompt is complete and well-scoped

Model-Specific References

Read the appropriate reference based on target model:

Target Model	Reference File
GPT-5.2, GPT-5.2-Codex	references/openai.md
Claude Opus 4.5, Sonnet 4	references/anthropic.md
Gemini 3 Pro, Gemini 3 Flash	references/google.md

Universal Handoff Components

Every handoff prompt should include:

Required

• Objective: Clear, specific goal
• Scope/Boundaries: What is and isn't in scope
• Output Format: Expected deliverable structure
• Constraints: What not to do, limitations

For Fresh Start Handoffs (add these)

• Project Context: Essential background
• Entry Points: Key files to read first
• Current State: What's done, what remains
• State Files: Progress tracking files to check

For Sub-Task Handoffs (add these)

• Dependencies: Files, APIs, or prior outputs needed
• Artifact References: Shared state or outputs
• Coordination Notes: How this task fits with parallel work

Quick Templates

Sub-Task Handoff (Universal)

<task_handoff target="[MODEL]">
<objective>[Specific, atomic goal]</objective>
<context>[Only what's needed for THIS task]</context>
<dependencies>[Files, APIs, prior outputs needed]</dependencies>
<scope>
  <include>[What to do]</include>
  <exclude>[What NOT to do]</exclude>
</scope>
<output>
  <format>[Structure of deliverable]</format>
  <location>[Where to save/return results]</location>
</output>
<coordination>[How this fits with parallel work]</coordination>
</task_handoff>

Fresh Start Handoff (Universal)

<fresh_context target="[MODEL]">
<project>
  <name>[Project name]</name>
  <overview>[1-2 sentence description]</overview>
  <entry_points>[Key files to read first]</entry_points>
</project>
<state>
  <completed>[What's done]</completed>
  <remaining>[What needs to be done]</remaining>
  <state_files>[progress.txt, tests.json, etc.]</state_files>
</state>
<task>
  <objective>[Specific goal]</objective>
  <success_criteria>[How to verify completion]</success_criteria>
</task>
<constraints>
  [Scope limits]
  [What not to do]
</constraints>
<output>
  <format>[Expected structure]</format>
  <verification>[How to validate results]</verification>
</output>
</fresh_context>

Model-Specific Adjustments

After generating the base prompt, apply these adjustments:

GPT-5.2/Codex

• Use CTCO framework (Context → Task → Constraints → Output)
• Add <reasoning_effort> tag (minimal/low/medium/high)
• For fresh handoffs, format as AGENTS.md

Claude Opus 4.5/Sonnet 4

• Avoid word "think" for Opus 4.5 (use consider/evaluate)
• Add explicit action mode (proactive vs conservative)
• Include parallel execution guidance
• Reference git for state tracking

Gemini 3 Pro/Flash

• Add thinking_level (LOW/MEDIUM/HIGH)
• Include anchoring phrase ("Based on the above...")
• Avoid broad negatives; be specific
• Note: keep temperature at 1.0

Best Practices

1. Minimize context - Include only what's essential for the task
2. Be explicit - State goals clearly; don't rely on inference
3. Scope tightly - Prevent overlap with parallel tasks
4. Include verification - How will the agent know it succeeded?
5. Reference artifacts - Point to shared state rather than duplicating
6. Match model style - Use patterns the target model responds to best