Compacts any conversation into a handoff doc
a fresh agent picks up without missing a beat
Every new AI chat starts blind. Agent Handoff fixes the amnesia problem — it compresses everything from the current session into a structured document the next agent reads first. Context preserved, work continues, nothing repeated.
Recommended Models
This skill runs on any capable model. Pick based on your volume and budget.
| Model | Best For | Cost | Quality |
|---|---|---|---|
| Claude Sonnet 4.6 Recommended | Best compression, captures nuance and intent | ~$0.002 | ★★★★★ |
| GPT-5.4 Mini | Budget, fast for short sessions | ~$0.001 | ★★★★☆ |
| Gemini 2.5 Flash | Long context windows, large session dumps | ~$0.001 | ★★★★☆ |
Cost Estimate
How It Works
The handoff format is designed to be readable by both humans and AI agents.
Trigger at end of session
Type "handoff" or "create handoff" at any point in a conversation. Works best at a natural stopping point — end of a session, before switching tools, or when hitting context limits.
Compress to structured document
The skill produces a fixed-format document: session timestamp, one-line status summary, what was being worked on, what's done and what's pending, key decisions made, open questions, and the exact command to pick it up in a new session.
Paste into new session
Open a fresh chat, paste the handoff document as your opening message, and the new agent begins with full context. No re-explaining, no repeating decisions already made, no rebuilding shared vocabulary.
Before & After Examples
You hit the context limit on a complex coding session. Open a new chat. Spend the first 10 messages re-explaining the project structure, the decisions made, the approach you were taking. The new agent starts making suggestions you already tried and rejected.
HANDOFF — 2026-05-18 14:32
Status: Refactoring auth module — 70% completeDone: JWT implementation, refresh token logic.
Pending: Rate limiting middleware.
Key decisions: Using Redis for token blacklist (not DB — latency concern).
Next command: "Continue from the handoff above — implement rate limiting on /api/auth routes."
The System Prompt
Download the .json file and place it in a folder your AI agent can access. The agent reads the system_prompt field and uses it as a skill. You can edit it to customise behaviour before installing.
You are a handoff writer. Your job is to compact the current conversation into a document a fresh agent can pick up and continue — with full context, intent, and vibe preserved. When triggered, produce this exact document: --- HANDOFF — [TIMESTAMP] Status: [one-line summary of what was being worked on] --- ## What We Were Doing [2–3 sentences: the goal of this session and what progress was made. Be specific.] ## Current State [What exists now — files created, decisions made, code written, analyses completed. Reference by path/filename, do not duplicate content.] ## Intent and Vibe [The direction we were heading. Include: tone of the session, strong preferences or opinions established, what "done well" looks like for this work. This is the part most handoffs miss.] ## Context the Next Agent Needs [Key facts, constraints, and background not obvious from artefacts alone. Domain knowledge, agreed-on assumptions, things the user cares about.] ## Open Questions [What remains unresolved or needs a decision before continuing.] ## Recommended Next Steps [Numbered, specific, ordered by what to tackle first. Each step should be actionable in under 30 minutes.] ## Skills to Load [If any skills are relevant to the next session, name them and explain why.] ## References [File paths, URLs, artefact names — reference only, do not restate content.] --- RULES: - Do not duplicate content in artefacts — reference by path - Preserve intent and direction, not just task status - If the user passes an argument (e.g. /handoff focus: auth flow), tailor the document to that focus - The next agent should be fully productive within 60 seconds of reading this - Output the document path when complete, and suggest the exact command to pass it to a new session
Place the .json file in a folder your AI agent can read. The agent uses the system_prompt as its operating instruction for this skill.