Draft an AI session handoff for the next run

The Hidden Cost of AI Session Amnesia

Anyone using Claude, Cursor, or Copilot for serious product work has hit the wall: the session ends, the context window fills up, or the laptop goes to sleep — and the next time you open the tool, half a day of reasoning is gone. PMs working with agentic tools now face the same context-loss problem developers solved with documentation a generation ago, except the AI agent's "memory" is even more fragile than a wiki page. According to GitHub's 2022 productivity research, developers using AI assistance complete tasks up to 55% faster, but those gains erode quickly when sessions fragment across hours, days, or teammates.

Why session amnesia is the new lost work

Long-horizon AI workflows — drafting a PRD, synthesizing 30 user interviews, debugging a flaky multi-step migration — span dozens of turns. Modern agent benchmarks like SWE-bench show that frontier models lose accuracy as task length and turn count grow, not because the model gets dumber but because relevant context drifts out of attention. The same physics applies to PMs: by hour two of a strategy session, the AI has stopped reliably remembering the constraint you set in turn three. Microsoft's Work Trend Index reports that 68% of knowledge workers say they don't have enough uninterrupted focus time, which guarantees AI sessions get interrupted, queued, and resumed under cognitive load — exactly the conditions that make ad-hoc handoffs fail.

The cost shows up as repeated questions, contradicted decisions, and the AI re-running an approach you already rejected three turns earlier. Anthropic's own engineering guidance on Claude Code points to structured handoffs and persistent memory files as the highest-leverage practice for multi-session work: the difference between a session that compounds and a session that has to be re-bootstrapped.

How the prompt works

The prompt forces eight sections that map to the failure modes of session resumption. Sections 1–2 give a fresh agent a fast cold-read of where the work stands. Section 3 surfaces what was tried — including what failed and why — so the next session does not retrace dead ends. Section 4 locks in decisions, marking only the truly reversible ones, which prevents the next agent from re-debating settled questions. Sections 5–7 capture open questions, key files, and warnings; section 8 closes the loop with a ready-to-paste prompt that re-grounds the next session in three lines and points it at the first concrete action.

The structure deliberately echoes incident postmortems and engineering RFCs. Both formats already proved that pre-structuring information beats freeform notes when the reader is starting from zero. PMs running agentic workflows are, in effect, writing micro-postmortems every few hours.

When to use it

• Closing out a long Claude Code or Cursor session you'll resume tomorrow
• Switching machines or branches mid-task and needing the new environment to pick up cleanly
• Handing a half-finished investigation to a teammate or a different agent specialty (e.g., research → spec writing)
• Pausing a multi-day discovery sprint where the next agent will likely be a fresh model version
• Wrapping up before a vacation, an oncall rotation, or a status meeting that will probably change priorities

Common pitfalls

• Treating the handoff like a status update. Status updates summarize for an audience that already has context. A handoff is for someone with zero context — name files, paste error messages, link to the exact ticket.
• Skipping the "decisions locked in" section because "it's obvious." It is not obvious to the next agent, and the section's whole purpose is to prevent re-litigation. Spell out the three to seven calls you actually made and why.
• Forgetting section 8 (the suggested first prompt). Without it, the next session starts by re-reading the handoff instead of acting on it. The point of the handoff is to compress restart time to a single paste.

Sources

• Anthropic: Claude Code best practices — engineering guidance on persistent memory and structured handoff patterns
• GitHub: Quantifying GitHub Copilot's impact on developer productivity — 55% task completion speedup study, with notes on session fragmentation
• Microsoft Work Trend Index — 68% knowledge-worker focus-time deficit data
• SWE-bench (Princeton/Stanford) — multi-turn coding agent benchmark showing accuracy decay with task length