Submit it when the employer asks, as Rootly does, and offer it when AI use is allowed but the brief is silent. Rootly says transcripts separated two candidates whose code was nearly identical. Before you send one, read it: tools such as Claude Code store file contents and command output in plain text, so remove secrets and anything that is not part of the take home.
Who asks for transcripts today?
In the sources we checked, Rootly is the only company that asks candidates to submit AI session transcripts. Since February 2026 its take home asks for them in the repository, next to the code and a video walkthrough under five minutes, and calls them a first class part of the submission.
Elsewhere the prompt trail is watched live instead of submitted. Hello Interview reports that Meta's AI enabled round puts the AI chat in a panel the interviewer can see, and Superhuman's technical round, according to LeadDev, has interviewers watch the candidate's chain of prompts.
What will a reviewer read in it?
Rootly's reviewers look at four things: how you broke the problem down, how quickly you caught and corrected the AI, whether your prompts show you understand the domain, and whether you iterated or went round in circles. The warning signs they describe are pasting the entire spec into one prompt, accepting output largely as it came, and asking for the same fix again and again.
Their example of two near identical submissions shows why it matters. The code alone would have been a coin flip; the transcripts showed one candidate catching the AI's mistake in retry handling and the other looping on fix it.
What is in a transcript file, and what should you check first?
More than the chat. Claude Code's documentation says each session file holds every message, tool call and tool result, and that the files are plain text: anything that passes through a tool, such as file contents, command output or pasted text, is written into the transcript. If a tool reads a .env file, the values in it end up there too.
So do the take home in a clean folder that contains only the take home, never with work credentials in reach. Before you submit, search the file for keys, tokens and anything from other projects. If you remove something, say so in the README; quietly editing a transcript is worse than sending a messy one.
How do you export one?
In Claude Code, run /export to save a readable plain text copy, or take the raw file from ~/.claude/projects/<project>/<session-id>.jsonl. Files older than 30 days are deleted by default, so export soon after you finish. Codex CLI keeps its local state under ~/.codex. Cursor writes agent transcripts as JSONL files which, according to a reply on Cursor's forum, leave out tool outputs. Our transcript locations page has the details.
Ask the employer which format they want. A readable export is easier for a person to review; the raw file carries timestamps and tool calls a reviewer may want.
Can a transcript hurt you?
Yes, and that is the point of asking for it. A transcript shows whether you read the code, whether you noticed when the AI went wrong, and whether your decisions were yours. If it shows a single giant prompt followed by repeated requests to fix it, the code will not save you.
That is also why it is worth practising before a real one counts. Rootly has not said how it handles edited transcripts or how consistently its reviewers agree, so the safest strategy is to work the way you would want to be seen working.
How can you practise for transcript review?
Do one of our practice rounds with your own agent, then paste the review prompt from the round's REVIEW.md into the same tool. It asks the agent to read your session and quote the moments that earned or cost you points on each rubric dimension. For plain counts, such as prompts, rejected edits and test runs, drop the file on the transcript stats tool, which reads it in your browser and uploads nothing.