Lionel Q.
← Field notes

I LEFT MY SUCCESSOR AN AI SECOND BRAIN, NOT A HANDOVER DOC.

Topic
Handover
Reading time
8 min

I had just sent my handover plan. Thirty-seven pages, built from chat threads, calendar invites, the wiki and call recordings. It was thorough, and every claim had a source.

One of the two CSMs who would inherit my accounts replied: "it's really long, tricky to understand actionable next steps."

Then a second line, the one I kept thinking about: "I really value your take and opinion."

That was the problem in one sentence. The document had the facts but not my judgment. It said what had happened on each account, not what I thought would happen next or what I would do in their place.

I had made my resignation official in mid-October, with a last day in mid-November. That left about three weeks to hand over roughly twenty enterprise accounts at a B2B SaaS data company, split across two CSMs. I had already spent part of that time on a document that answered the wrong question.

  1. 01

    Why a Solutions Engineer handover doc fails

    Rep transitions are expensive, and researchers have put numbers on it. A Journal of Marketing study of a Fortune 500 B2B firm (opens in a new tab) found that sales-rep transitions caused a 13.2 to 17.6% annual sales loss. One detail stuck with me: replacements with similar industry experience limited the loss better than the "stars" did.

    A second Journal of Marketing study (opens in a new tab) put the cost of relationship disruption at up to 17.6% of customer revenue, with resale revenue down 28.1%. It also notes that sales turnover runs at 20% a year or more.

    Customer-facing people move on often. ChurnZero's 2025 research (opens in a new tab) found that 58% of CSMs would leave for their ideal role elsewhere. And much of what they know is theirs alone: 2018 vendor research from Panopto (opens in a new tab) estimated that 42% of institutional knowledge is unique to the employee's role.

    Michael Polanyi named the underlying problem in 1966: "we can know more than we can tell." Nonaka's SECI model (opens in a new tab) makes it practical. A typical handover doc is Combination: taking explicit knowledge (tickets, notes, emails) and rearranging it into more explicit knowledge. The hard step is Externalization, turning tacit knowledge into something written: who the real champion is, which promise was a stretch, which account is quietly drifting.

    Good practice exists. GitLab's public handbook has a CSM-to-CSM handoff checklist (opens in a new tab). But it is a checklist of tasks, and judgment about the customer is not a task you can tick.

    Customers feel the gap too. Bendapudi and Leone wrote in HBR in 2001 (opens in a new tab) that customers fear depending on one contact and worry that "a replacement won't be as good." Their fix: announce the successor early and have the departing person make the introduction.

    My 37 pages did Combination well. They skipped Externalization entirely.

  2. 02

    What I built: a handover knowledge base with my judgment in it

    First, I cut it down. One page per CSM, one table: account, interim owner, scenario if a new hire arrives, hot topics, my sentiment.

    Second, I got interviewed. The sentiment column was empty, and I was not going to write twenty honest paragraphs with three weeks left. So I had Claude interview me, one account per question, with quick multiple-choice panels: solid, demanding, to watch, at risk.

    It was fast enough to cover twenty accounts in minutes, and because each question was narrow, the answers came out blunt. A few, anonymised:

    • "The champion is a real strength: go step by step, finish the identity project first."
    • "Main risk is they build in-house tech that replaces us."
    • "At risk: hard to reach, little visibility."

    None of that was in the 37 pages.

    Third, I turned it into a knowledge base. Parallel AI agents researched each account across chat, email, call recordings and CRM-like admin data. The output went into Google Drive, one folder per account, four documents each:

    HANDOVER/
    ├── START_HERE           index: how to read the knowledge base
    ├── FACTS_TO_VERIFY      claims a human should confirm
    └── <Account>/
        ├── README           who, contacts, channels, environments, stack
        ├── HISTORY          dated, sourced chronology
        ├── OPEN_ITEMS       what's live, who owns it, what the customer
        │                    is waiting for, what I promised in my name
        └── OWNER_NOTES      my sentiment verbatim, what to do,
                             what not to do

    About 80 documents in total. Three rules held across all of them:

    1. Every fact carries a source link.
    2. If something was not found, the doc says "not found" instead of guessing.
    3. My voice is labelled as mine. Everything else is marked "(from sources)".

    The research surprised me. One account lived almost entirely in email, not chat, and the first pack missed most of it. The agents also surfaced things I had not noticed: jobs failing silently, a champion laid off the day before, and a commitment made by a colleague that engineering couldn't meet.

  3. 03

    Turning the knowledge base into a Claude plugin

    Tiago Forte puts it this way (opens in a new tab): "Note-taking is like time travel—you are sending packets of knowledge through time to your future self." In a handover, the future self is someone else, and they won't read 80 documents. They will ask questions.

    So I wrapped the knowledge base in a Claude plugin. Anthropic says building an Agent Skill (opens in a new tab) is like "putting together an onboarding guide for a new hire," which is close to literally the job. Plugins (opens in a new tab) "bundle any skills, connectors, slash commands, and sub-agents together," so colleagues install one thing and get all of it.

    Colleagues ask it things like:

    • "What's open on X?"
    • "Prep me for my X call."
    • "Which accounts are at risk?"
    • "Log this on X."

    Here is what an answer looks like, anonymised and trimmed:

    Q: Brief me on Account A before I take it over. Anything I shouldn't do?

    A: The previous owner's view, in his words: "The champion is a real strength: go step by step, finish the identity project first." (Source: Account A / OWNER_NOTES)

    So close the identity project before starting anything new with this customer.

    Commercial figures: not found in the connected sources. Confirm them before your intro call. (See FACTS_TO_VERIFY)

    The answer carries my opinion, labelled as mine, next to sourced facts, and it says plainly what it couldn't find.

    "Log this on X" matters most over time. It adds a dated, sourced entry to HISTORY or OPEN_ITEMS, so the knowledge base keeps going after I've left.

    Once it worked for my own portfolio, I generalised it. se-handover is an open-source, company-agnostic Claude plugin with nine skills:

    StageSkillWhat it does
    Planhandover-planAccount inventory, owners, timeline
    Capturesentiment-interviewAccount-by-account interview with quick panels
    Captureaccount-packSourced four-document pack per account
    Publishknowledge-base-publishFolder per account, START_HERE, FACTS_TO_VERIFY
    Publishowner-one-pagerOne table per CSM
    Useaccount-brief"What's open on X?"
    Usecall-prepBrief before a customer call
    Useportfolio-view"Which accounts are at risk?"
    Maintainlog-update"Log this on X"

    Tools are referenced by category (~~chat, ~~email, ~~CRM, ~~cloud storage), so any company can map it to its own stack.

  4. 04

    Limits: what an AI handover can't do

    It only knows what was written down somewhere. If a decision happened in a hallway, it isn't there.

    The CRM wasn't always connected for me, so commercial numbers in the packs need confirming. That is what FACTS_TO_VERIFY is for.

    Packs go stale unless people log updates. A knowledge base nobody writes to is just a slower way to be out of date.

    The AI can be wrong. A Stanford study of commercial AI legal research tools (opens in a new tab) found they hallucinated 17 to 33% of the time, and concluded that users "must continue to verify that key propositions are accurately supported by citations." That is why every line has a source, and why "not found" beats a confident guess.

    And it does not replace the human introduction. Bendapudi and Leone's advice still holds: the departing person should make the intro call. The plugin makes that call better prepared, but you still have to make it.

  5. 05

    Try the Claude plugin before your last day

    Install it:

    /plugin install se-handover --marketplace lionelquellery/se-handover-plugin

    Then, in this order:

    1. Run handover-plan to get the inventory and owners.
    2. Run sentiment-interview while your opinions are still fresh.
    3. Run account-pack, then knowledge-base-publish.
    4. Give your successors the plugin and show them account-brief and log-update.

    Fork it before your own last day. (opens in a new tab) If it helps, star the repo. If your handover had gaps it didn't cover, open an issue and tell me what they were.

  6. 06

    Sources

← All field notes