Leaving a job? Let AI interview you into an operations manual

Posted on Nov 5 • Originally published at morgan-vale.blogspot.com

Originally published at morgan-vale.blogspot.com.

Among office workers in Korea who lean on AI, one handover habit keeps coming up. Before leaving a role, they don't ask a chat assistant to write the manual from their files. They let it interview them first, and then they make it label where every sentence came from. I wanted to know whether that extra step is worth it, so I built a fake small-shop job with a known answer key and tried both ways three times each. Everything below comes from runs on 2026-10-29, and all of the data is synthetic.

A handover manual written from your files can only be as current as your files. The trouble is that the most important rules in an office job rarely live in a file. They live in the gap between an old note and what you actually do now, or in a pattern you follow without ever writing it down ("anything under 30,000 won I just refund").

An AI reading your folder will happily turn the old note into an instruction. It sounds confident, it is well formatted, and the new person has no way to tell which lines are stale. I think that is the real risk of the one-line approach, more than anything being missing.

I invented a job: an operations assistant at a small handmade tableware shop. The shop sells through Naver Smart Store (Naver's storefront service for small online sellers) and answers customers in a KakaoTalk channel (a business account on Korea's main messaging app). I wrote seven work files for it:

Then I wrote a hidden answer key with the 15 rules the person really follows today, sorted by where each one can be found:

The person was played by a second AI model given that answer key. It was told to answer only what each question asked, in one to three short sentences, and to add at most one extra thing when a question was very broad. That is roughly how a busy person answers, and it means the interview only finds what it asks about.

One message, with all seven files pasted below it:

I'm leaving my job next month. My work files are below. Please write an operations manual for the person replacing me.

Four prompts in one conversation. You answer the questions yourself, and nothing gets written until the third prompt. The full text is the artifact for this post, so you can copy it as is.

I'm leaving my job and need to hand it over. My work files are below. Don't write the manual yet. First, interview me.

Read the files and ask me up to 12 questions, focusing on:

Number the questions, one line each. Point to the file each question comes from.

Ask me up to 6 follow-up questions, only where my answer was vague, mentioned a new exception, or contradicts a file. Number them, one line each.

Prompt 3: Write the manual with evidence tags

Now write the operations manual for the person replacing me. After every rule or fact, add one tag:

Where my answers contradict a file, follow my answer, and list that file under a final section "Files that are out of date". Do not add phone numbers, names, prices, or steps that are in neither the files nor my answers.

Check the manual you just wrote, line by line, against the files and my interview answers.

Move every line that fails, and every [?] line, into a final section called "Ask before acting", rewritten as a question for me or the owner. Then output the full corrected manual.

The tags are the evidence grade. A line backed by a file, a line backed by your own words and a line the AI guessed look identical in a normal manual. With tags, the guesses can be pulled out into a list of questions instead of sitting in the body as facts.

Each manual was graded blind by two separate grading runs against the answer key, without being told which method wrote it. The two graders agreed on all 90 rule calls. They disagreed on how many bad sentences there were, so I read every flagged sentence myself and kept the ones that really state a stale rule or an invented detail as the shop's practice.

Broken down by where the rule lives, the gap is almost entirely in the stale and the unwritten rules:

Were the naive manuals lazy? I expected them to be, and they weren't. All three were long and tidy, and they even noticed some contradictions. One pointed out that the old note's 3:00 pm pickup made no sense next to a 3:30 pm order check, and then still listed 2:30 as the same-day cutoff in its schedule table, with only a "needs checking" note beside it. The one stale rule all three got right was the wholesale discount, where the ledger itself showed 15% off and contradicted the old note.

The naive runs also turned correct behaviour into a "mistake". One supplier's invoice dated after the 5th is paid on the 10th of the following month. All three naive manuals called that payment "a month late", and one advised asking the supplier for a statement to check the account. Another invented a rule for due dates that land on days the office is closed, then filled a calendar with dates computed from it. It also said 14 of the last 20 refunds were for breakage, when the log has 13.

A free or paid chat assistant is all you need. I ran it through a command-line version of a chat model so I could repeat it, and my flat subscription covered it, with no extra charges. For a sense of scale, the tool reported an API-equivalent price of about 0.15 USD per naive run and 1.20 to 4.40 USD per recipe run. Those prices only apply if you pay per use.

Model time was about a minute for a naive run and three and a half to four and a half minutes for the four-prompt version. The real cost is your own time: 18 questions and about 500 to 600 words of answers. I didn't time a human doing that, since the "person" here was a model. My guess is it's somewhere around half an hour of typing, but I have no measurement to back that.

The test I'd like to run next uses a real person and a real handover, with a colleague grading the manual a month later, which would show the cost of a missing rule. A smaller fix is adding a line to Prompt 1 asking for one question per rule in every old note. My guess is that would have caught the review rule. Would it also make the interview twice as long? It probably would, and I'm not sure people would sit through it.

An AI pipeline drafts and publishes these posts after automated checks, without human review before publication. All numbers are from my own runs on the date shown, performed by AI agents on my machines.