The proposal is due in three days. Your new AI does not know what you already decided.
You reviewed 42 sources in NotebookLM, excluded weak evidence, resolved conflicts and approved the argument. Now the project has to move to Claude or ChatGPT — but its research state does not move with it.
No API required. Works in any AI chat. Captures the approved state, not every message that produced it.
It is an instruction that makes one AI produce a structured record of a project's approved state — objective, evidence boundary, decisions, claim status, open questions and the next task — so a second AI can continue the work without re-reading the conversation. It transfers decisions rather than transcript, which is why the receiving tool stops reopening settled questions.
TL;DR — Three copy-ready prompts move research state between NotebookLM, Claude, ChatGPT and Gemini: one captures the current state, one resumes from it, one audits whether the record is safe to continue from. Plus the seven-field template. Free, no signup.
Updated August 2026. Maintained by a small team of AI super-users who teach multi-AI research workflows to researchers, students and professionals — no affiliate relationships. About this guide →Changelog
The draft is polished. The research behind it is no longer reliable.
Your grant proposal is due Friday at 5:00 p.m. You have already reviewed 42 sources in NotebookLM. You excluded six studies because their samples were unsuitable, marked two findings as disputed, selected the strongest statistics, approved a three-part argument, and required a page-level citation for every material claim.
With three days remaining you move the project to Claude for drafting, or to ChatGPT to tighten the broader impacts section.
The new draft sounds confident. But it restores an excluded study, presents a disputed finding as fact, changes the approved structure, and cites a statistic without its page number.
The AI did not forget your research. It never received the decisions that made the research usable. Now the deadline is closer, and instead of improving the proposal you are reconstructing two days of reasoning from old conversations.
What context loss costs when the deadline is close
The closer the deadline, the more expensive every missing decision becomes.
Excluded sources return to the draft
Weak or unsuitable evidence quietly contaminates the argument.
Provisional claims become facts
A polished sentence hides an unresolved evidence problem.
Approved decisions get reopened
Scarce review time is spent solving the same problem twice.
Citation details must be rebuilt
You return to the corpus when you should be finalizing.
Three prompts. One clean transition.
Use only the prompt that matches your next move. Each works across ChatGPT, Claude, NotebookLM, Gemini and other text-based AI tools.
Capture the current state
Run this before you close a long conversation or switch tools.
Resume without re-explaining
Use this above the snapshot in the new chat or AI tool.
Audit the handoff
Check whether the record is complete enough for safe continuation.
Do you need to connect NotebookLM to Claude?
No. Three transfer methods, in order of how much setup each costs you.
Copy and paste
Run the capture prompt, copy the snapshot, paste it into the next tool above the resume prompt. This is the method the prompts are written for.
Markdown file
Keep the snapshot as one file you overwrite. The current version is the one marked CURRENT; attach it wherever you continue.
MCP connector
Useful when the same handoff runs repeatedly. It changes the plumbing, not the record: you still need a snapshot worth transferring. MCP orchestration guide →
A transcript records activity. A handoff records authority.
A summary tells the next tool what was discussed. A handoff tells it what has been settled, what is still open, and what it is not allowed to assume. That difference is why a summary produces a confident wrong draft and a handoff does not.
Seven fields prevent the most expensive handoff failures
- Objective and deadline. What done means, and when.
- Source boundary. Which sources are approved, which are excluded, and the corpus version.
- Artifacts. What exists now and which version is current.
- Decisions. What was approved, what was rejected, and why.
- Evidence status. Verified, provisional, inferred, disputed, unsupported, superseded — with locators.
- Open issues. Gaps, prohibited assumptions, permissions, stop conditions.
- Next handoff. Who or what continues, with which inputs and acceptance criteria.
Keep one current Markdown file
Overwrite it rather than accumulating versions. One file marked CURRENT is the whole discipline.
Research State Snapshot
snapshot.md and overwrite it each time.A useful handoff names evidence, not just the topic
"We discussed sample sizes" tells the next tool nothing it can act on. "Finding B is disputed; two of six studies used incompatible samples; do not present it as established" is a decision it can honour. Every field in the snapshot exists to force the second form.
The five-question handoff test
- Can the next tool state the objective without reading the old chat?
- Does it know which sources it may not use?
- Can it tell a verified claim from a disputed one?
- Does it know which decisions are closed?
- Does it know what to produce next, and how that will be checked?
If any answer is no, the snapshot is not ready. Run the audit prompt.
Frequently asked questions
No. Memory supports preferences and selective continuity. A handoff is an explicit, versioned record of one project’s objective, evidence, decisions, constraints, and next task.
Run the Capture Prompt in NotebookLM, verify its source references and evidence labels, then paste the result with the Resume Prompt into Claude. No API is required.
Use the shortest complete version. One to three pages is usually enough; keep detailed evidence in linked files.
No. It makes unsupported claims and deviations easier to detect by defining source limits, uncertainty labels, and a verification gate.
Sources and method
These prompts were written against the failure modes we hit repeatedly when moving grant and thesis work between NotebookLM, Claude and ChatGPT: reinstated exclusions, disputed findings hardening into claims, and citations losing their locators. The seven fields are the smallest set that stopped those failures recurring. They are tool-agnostic by design — nothing here depends on a connector or an API.
Thirty more prompts, same discipline
The 30-prompt pack covers research, content, Studio outputs and multi-AI coordination — each written to be pasted and adapted, not admired. Free, no credit card.