Automation & Agents

Recovering an Interrupted AI Task: Rebuild the Timeline from Your Transcript

An interruption destroys the chat window, not the record. Tools like Claude Code save a full transcript of every session on your disk, so recovery means telling the AI to verify, not to remember.

10/02/2026 ｜ 3 min read ｜ A recovery prompt you can save and paste next time — works in 10–20 minutes

The one idea that matters

An interruption destroys the chat window, not the record. AI workbenches keep a full transcript of every session on your own machine. So the right recovery is not to ask the AI to “remember” (with no record it will make things up), but to make it verify: read the files, rebuild a timeline, show it to you, and only then continue.

Step 1 — Reopen the tool and paste the recovery prompt

Reopen Claude Code in the same project folder and paste this (edit as needed):

My computer just lost power (or crashed / hit a usage cap). I remember we were working on <one-line task description>. Please recover. Rule: do not guess from impression. Verify in this order:

1. When the project notes / memory file were last updated
2. File modification times in the working folders (when did activity stop)
3. The previous session’s transcript, to find the exact breakpoint (which step, what the last instruction was)
4. The previous session’s temp folder, for salvageable partial output Show me a factual timeline and your breakpoint conclusion first, then continue.

Step 2 — Know where the transcript lives

• Windows: C:\Users\<you>\.claude\projects\<project-folder>\<session-id>.jsonl
• Mac: ~/.claude/projects/<project-folder>/<session-id>.jsonl
• With many files, sort by last-modified time; the newest is usually the interrupted session.
• Each line is one JSON record: your instructions, the AI’s replies and plans, sub-task reports, exact timestamps.

If the AI can’t find it, ask:

I use <tool> on <Windows/Mac>. Tell me where this tool stores its conversation logs, then list the logs updated in the last three days. One step at a time.

Step 3 — Check the scratch folder for partial output

Claude Code’s session scratchpad lives under the system Temp folder (on Windows roughly C:\Users\<you>\AppData\Local\Temp\claude\<project>\<session-id>\scratchpad\). Ask the AI:

Check the previous session’s scratchpad and tasks output folders and list every file with its size. Interrupted sub-tasks may have left complete output.

Step 4 — Confirm the timeline yourself, then release it

Spend 30 seconds on the AI’s timeline: is the breakpoint plausible? Does each “done” item have a file to back it up? Reply “continue” only when it checks out. That 30 seconds is the last gate against guesses passing as facts.

Three traps you will hit

Trap 1 — Treating a “failed” sub-task as having produced nothing, and redoing a whole round. A sub-agent killed by a usage cap may have finished writing its output a second before it died. In one case a sub-task marked failed had a complete 30KB research report sitting in scratch. Check for written files before deciding to re-dispatch.

Trap 2 — Leaving intermediate output in scratch until the system cleans it up. Scratch is not a safe. After each stage of a long task, have the AI save intermediate output to a permanent folder. Recovery is insurance; persisting early is the real fix.

Trap 3 — Letting the AI continue “from memory,” so it confidently invents progress that never happened. Always include “verify first, show the timeline, wait for my go-ahead.”

When you’re stuck, ask the AI like this

I’m a beginner using <tool> on <OS>. I’m stuck on step <N> of recovering an interrupted AI task. What I actually see / the error text is: <paste it>. Explain in the plainest words, give me one step at a time, and wait for my result before the next.

In one line

A power cut isn’t the disaster; the record is on your disk. Save the four checks (notes, timestamps, transcript, scratch folder) and paste them next time you’re interrupted.