AI Project Handover: Keep Decisions Clear Across Chats

An AI-assisted project can begin with a clear brief and acquire dozens of useful decisions: the audience changes, a draft is approved, a proposed feature is postponed and a source turns out to be unreliable. When work moves into a new chat, recreating those decisions can become a project of its own.

A short, maintained handover gives the next session a usable starting point. It records the current goal, approved decisions, source locations, unresolved questions and the next action. You can inspect it before relying on it, share the relevant parts with a collaborator and correct it when circumstances change.

The aim is to carry forward enough verified project state to continue the work accurately. For creators and small businesses, that can mean preserving a book’s approved structure, a product’s agreed scope or a video’s chosen script without replaying every abandoned suggestion.

Why a long history still needs careful handling

In Lost in the Middle, published in 2024, Nelson F. Liu and colleagues tested language models on document questions and information retrieval. Performance varied with the position of relevant information in long inputs. This finding concerns the models and tasks studied; it does not establish how every current chatbot handles every project.

Anthropic’s engineering guidance of 29 September 2025 describes approaches including compressing conversation history and keeping structured notes for later use. It also warns that excessive compression can discard details that subsequently become important. These are engineering practices, rather than proof that one particular handover template will improve every user’s results.

The practical lesson is to preserve the decisions that affect the next task and retain a route back to the underlying evidence. A readable summary and an accessible source archive serve complementary purposes.

Record six parts of the current project

Use a structure that someone unfamiliar with the latest discussion could follow. The following six-part note is an applied editorial framework:

• Goal: the result being pursued, the intended audience and the current deliverable.
• Approved decisions: what has actually been agreed, with a brief reason where it affects future work.
• Current materials: the relevant files, versions, links and the location of the latest approved draft.
• Open questions: unresolved choices, missing information and assumptions still awaiting evidence.
• Boundaries: requirements to preserve, excluded work and actions requiring separate approval.
• Next action: one concrete step, together with the evidence or check needed to complete it.

Keep proposals visibly separate from approvals. A suggestion that sounded promising during brainstorming should not quietly become a confirmed product feature. Likewise, a completed draft should not be labelled published unless publication has been checked.

This extends the continuity principle in our article on writing a restart note. A restart cue helps you resume at the right point; a project handover adds the decisions and materials another working session needs.

A worked example for an independent creator

Consider a fictional creator preparing a beginner’s guide to recording short educational videos. A useful handover might read as follows:

Goal: Finish a practical guide for people recording their first educational video on a phone. The current deliverable is an edited manuscript, not a recorded course.

Approved decisions: Use British English, explain technical terms and organise the guide around planning, recording and reviewing. A chapter about paid advertising was postponed because it falls outside the beginner’s task.

Current materials: The approved outline is in Video_Guide_Outline_v3.docx. The working manuscript is Video_Guide_Draft_v5.docx. Both are in the project’s Manuscript folder. The earlier fourth draft is superseded.

Open questions: The accessibility section needs source checking. No decision has been made about adding printable worksheets. The cover concept remains a proposal.

Boundaries: Preserve the approved chapter order. Keep examples hypothetical unless their details have been verified. Editing the manuscript does not authorise publishing it.

Next action: Open the fifth draft and check the accessibility section against the research notes. Report unsupported claims and propose revisions before proceeding to the next chapter.

The filenames and project are illustrative. Their value lies in the distinctions: approved versus proposed, current versus superseded, and completed versus still requiring work.

Begin the next chat with an access check

Supply the handover and make the relevant sources available through the tools or attachments the session supports. A filename in a note identifies a source; it does not demonstrate that the next session can open that source.

Ask the assistant to state the current deliverable, identify the materials it can actually inspect and flag anything missing before it works on the next action. Check that response against the handover yourself. An assistant repeating a filename has not necessarily read the file.

Google’s prompt-design documentation, updated on 13 October 2026, recommends clear instructions and supplying the information needed for the task. It also treats prompt design as iterative. Applied here, that means stating the next task explicitly and checking whether the supplied context was sufficient.

Maintain the note when the project changes

Update the handover after a meaningful decision, a completed milestone or a change in the working files. Record a version or date so you can distinguish the latest note from an earlier snapshot. There is little benefit in maintaining an elaborate record nobody can identify as current.

If you ask AI to prepare the update, review the proposed changes against the source material. Check names, numbers, file versions and approval status. Preserve an important reason for a decision when losing it would invite the same unsuccessful approach later.

Keep enough history to understand a reversal. For example, record that the worksheet proposal was approved after a review, rather than leaving contradictory instructions in two places. When sources disagree, identify the conflict and resolve it explicitly.

Keep the evidence reachable

A handover should point to the work rather than replace it. Store longer research notes, drafts and supporting records where they can be retrieved. Our creative-project source ledger describes how to connect assets, versions and permissions to a finished piece.

When moving material between services, include only information appropriate to share there. Preserve necessary access boundaries alongside the project requirements. Review time-sensitive facts before reusing them: a previously accurate price, deadline or product description may need checking again.

Start with one active project. Write its six-part handover, use it at the next transition and note what information was missing or misleading. Revise the structure around those observations.

Good continuity becomes visible in the work: the right version opens, an approved decision survives, an unresolved question stays unresolved and the next action begins from evidence.

Sources and further reading

• Liu, N. F. et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173.
• Anthropic (29 September 2025). Effective context engineering for AI agents.
• Google AI for Developers. Prompt design strategies, updated 13 October 2026.