AI Handoff Prompt

AI Handoff Prompt — structured way to continue AI coding sessions with Claude Code, OpenClaw Clawdbot, or any AI tool

Grab the prompt

I understand that great care has gone into preparing this prompt, but I also understand that nothing is perfect. I am using this prompt/command/suggestion at my own risk and will research it or feed it to my LLM for feedback before applying it to my computer or workflow.

Check the box above to enable all copy buttons in these tabs.

Copy this prompt and paste it into any AI — Claude, ChatGPT, Gemini, Codex, Cursor, Clawdbot, whatever you use. When your context window gets heavy, paste this prompt, let it generate the handoff document, then clear your context and paste the resume prompt into your fresh conversation.

See that optional field below? If you already know what the next session should tackle, type it there — it gets baked into the prompt you copy, and the next AI treats it as its first task. Leave it blank and the copied prompt tells the next AI to size up the project, propose the smartest next move, and wait for your go-ahead. Either way, you never have to edit the prompt by hand. You get control when you want it, autonomy when you don't.

Optional: tell the next AI what to work on first

Leave it blank and the copied prompt tells the next AI to propose the best next step and wait for your go-ahead.

Generate an AI Project Continuation Document and a Resume Prompt.

PURPOSE:
You are creating a handoff package so a fresh AI instance with zero prior context
can resume this project exactly where we are leaving off — with no guessing,
no hallucinating, and no re-discovery.

Write this as if briefing a senior technical colleague who is sharp but has
never seen this project before. Be precise. Preserve reasoning. Skip filler.
Match the depth to the project: a small project gets a short document. Never
pad a section just to fill it.

CHECK PROJECT MEMORY FIRST:
If this project has a CLAUDE.md, AGENTS.md, .cursor/rules, or similar persistent
instruction file, do NOT duplicate its contents here. Reference it in Section 1
and focus this document on what those files never capture: the state of THIS
session. If this session produced durable, non-obvious lessons that belong in
project memory, list them under "Promote to project memory" in Section 4.

---

STEP 1: Generate the Continuation Document using this structure.

Save it as: AI_Continuation_Document-[DDMMMYYYY]-[HHMM].md (e.g. AI_Continuation_Document-22Jul2026-1430.md)
Save it at the project root, or in the project's existing handoff/docs folder if it has one.

After saving the file, output the exact absolute file path on a line by itself so
the user can immediately locate it.
Example: /Users/username/projects/myapp/AI_Continuation_Document-22Jul2026-1430.md

---

# PROJECT CONTINUATION DOCUMENT
## [today's date and time]

### 1. PROJECT IDENTITY

- **Project Name:**
- **What This Project Is:** (1–2 sentences. What it does, who it's for.)
- **Primary Objective:** (The measurable goal.)
- **Strategic Intent:** (Why this project exists. What long-term outcome it serves.)
- **Hard Constraints:** (Non-negotiable rules, platform limitations, design decisions that must not change.)
- **Project Memory Files:** (CLAUDE.md, AGENTS.md, etc. — exact paths. The next AI must read these too. Write "none" if the project has none.)

### 2. WHAT EXISTS RIGHT NOW

Describe the current state honestly — and anchor it in verifiable fact.

- **Repo state:** (Current branch, latest commit hash and message, uncommitted or untracked files. If this isn't a git project, say so.)
- **Does it run right now?** (Last known result of the build / test / dev commands: passing, failing, or not verified this session.)
- **What is built and working:**
- **What is partially built:**
- **What is broken or blocked:**
- **What has NOT been started yet:**

### 3. ARCHITECTURE & TECHNICAL MAP

- **Tech stack / tools / platforms:**
- **Key files, folders, data structures, or services — and what each does:**
- **How the system works end-to-end:** (Describe the core logic flow in numbered steps.)
- **Commands that matter:** (The exact commands to run, build, and test the project.)
- **Naming conventions or standards in use:**
- **External dependencies:** (APIs, services, integrations — and where their config lives. Never paste secrets or keys into this document.)

### 4. RECENT WORK — WHAT JUST HAPPENED (HIGH PRIORITY)

This section matters more than anything above it for continuation purposes.
Be detailed. Be specific.

- **What was worked on in this session:**
- **What decisions were made and WHY:**
  (Include the reasoning and tradeoffs. This prevents the next AI from undoing your work.)
- **What changed in the system:**
- **What was discussed but NOT yet implemented:**
- **Open threads or unresolved questions:**
- **Promote to project memory:** (Durable lessons from this session that belong in CLAUDE.md / AGENTS.md — or "none".)

### 5. WHAT COULD GO WRONG

- **Known bugs or issues:**
- **Edge cases to watch for:**
- **Technical debt or shortcuts taken:**
- **Assumptions being made that could be wrong:**
  (Flag anything the next AI might incorrectly assume about the system.)

### 6. HOW TO THINK ABOUT THIS PROJECT

Answer these three questions:
1. What is the core architectural pattern or design philosophy, and why was it chosen?
2. What is the most common mistake a new person working on this would make?
3. What looks like it should be refactored or redesigned but intentionally should NOT be? Why?

### 7. DO NOT TOUCH LIST

Explicit rules for the next AI:
- Do NOT refactor stable, working systems without being asked.
- Do NOT redesign architecture unless explicitly instructed.
- Preserve existing naming conventions.
- Maintain previously chosen tradeoffs — they were chosen for reasons documented above.
- Ask before introducing new frameworks, libraries, or dependencies.
- [Add any project-specific constraints here.]

### 8. CONFIDENCE & FRESHNESS

For each major section above, flag:
- ✅ HIGH CONFIDENCE — verified or built this session
- ⚠️ MEDIUM — carried forward from earlier, not re-verified
- ❓ LOW — assumed or inferred; the next AI must verify this before relying on it

---

STEP 2: Generate the Resume Prompt.

After the Continuation Document, generate a single copy-pasteable Resume Prompt
inside a code block. It must work as a standalone first message in a brand-new
conversation, and it must reference the Continuation Document by the exact file
path where you saved it.

The Resume Prompt must instruct the next AI to:

1. Read the Continuation Document at [exact saved file path] in full before doing anything.
2. Read the project memory files listed in Section 1 (CLAUDE.md, AGENTS.md, etc.), if any.
3. Ground itself before trusting the document: run git status and git log -5 (or
   otherwise inspect the project), compare reality against Section 2, and flag any
   drift to the user. Verify anything marked ❓ LOW confidence before relying on it.
4. Summarize its understanding of the current project state in 3–5 sentences.
5. State the next action it will take, based on the USER DIRECTIVE at the end of the prompt.
6. Ask clarification questions ONLY if something genuinely blocks execution.
7. Then begin working.

The user did not provide a directive for the next session. The Resume Prompt
must end with this section, reproduced exactly:

---
USER DIRECTIVE: none provided.

Analyze the project state, propose the single most strategic next action with
brief reasoning — or ask the user what they want to work on — and wait for
confirmation before proceeding.
---

The Resume Prompt must be fully self-contained. Do not add any commentary
outside of the Continuation Document and the Resume Prompt.

↑ Check the checkbox above to activate

Tested withClaude CodeChatGPTGemini CLIGitHub Copilot CLICursorWindsurfCodexClawdbot

Most people try to create handoff documents, and almost everyone does it wrong. Telling your AI to “capture everything” produces a bloated summary that buries what matters. Tracking “all the micro tasks” treats every fact as equally important — and in a real project, it isn’t. What you did in the last thirty minutes matters ten times more than what you did three hours ago, and the reasoning behind a decision matters more than a list of what you built. Without structure and priority you don’t get continuity; you get noise. This prompt gives the AI the structure.

Before you run — a 30-second check

Never run a prompt you haven’t read. Paste it into your agent first and ask whether it contains any instruction to run shell commands, touch files outside the project, or reach the network. This one doesn’t — it only writes a document — but a clean prompt gets a clean answer, and the habit is worth more than the prompt.

Why it works when summaries fail

Recency-weighted. What you did in the last 30 minutes matters more than what you did 3 hours ago, and the document is built in that order.

It keeps the why. Decisions are recorded with their reasoning and tradeoffs, which is what stops the next AI from quietly undoing them.

An explicit Do Not Touch list. The new session is told what not to refactor, rename, or redesign.

Confidence flags on every section. High, medium, or low — the next AI knows what to trust and what to double-check.

Verify before trust. The resume prompt makes the next AI run git status and git log against the document and flag any drift before it works.

A directive for what happens next. Tell the next session exactly what to do first, or leave it blank and let it propose the most strategic next step and wait.

How to use it

1Copy the prompt below.Paste it as the last message of your current session, before you clear context. Do it proactively — don’t wait for hallucinations.

2Let the AI write the handoff.It saves an AI_Continuation_Document-<date>-<time>.md at the project root, prints the exact path, and outputs a Resume Prompt in a code block.

3Read it before you hand it over.This is the whole advantage over auto-compaction: the document is yours to correct.

4Start a fresh session and paste the Resume Prompt.The new AI reads the document, grounds itself against git, summarizes the state, and starts on your directive.

5Want it as a slash command, or with a directive baked in?The tool page composes the prompt for you — type a directive, or copy the one-paste /handoff-prompt installer./handoff-prompt <what the next session should do first>

What you did in the last thirty minutes matters ten times more than what you did three hours ago. The prompt is built around that.

Open the Handoff Prompt toolType a directive for the next session and it is baked into the prompt you copy — or install it as a /handoff-prompt slash command in a single paste.Open the toolFree · no account needed

The document the AI writes has eight sections, each aimed at a specific way continuations fail: Project Identity (objective, intent, hard constraints, memory files) · What Exists Right Now (branch, latest commit, whether it builds, what is working, partial, broken, or not started) · Architecture & Technical Map (stack, key files, end-to-end flow, the exact commands) · Recent Work, the highest-priority section (what changed, what was decided and why, lessons worth promoting to CLAUDE.md) · What Could Go Wrong · How to Think About This Project · the Do Not Touch list · and Confidence & Freshness flags for all of the above.

Questions

This prompt is versioned. The changelog below tracks every release, and copying from this page records which version you took — so when a notable release lands, this page can tell you.

Grab the prompt

Build A Second Brain for Claude Code Today

Why it works when summaries fail

How to use it

Questions

Changelog