AI AgentIntermediate13 min read

Evidence, not vibes: how to check a Bot's work

The errors Bots make most, asking for proof, spot-checking three items well, matching effort to risk, writing the check into the brief, stop rules, and a list of guardrails and good habits.

Evidence, not vibes: how to check a Bot's work

Evidence, Not Vibes: How to Check a Bot's Work

A Bot can hand you something that looks finished, reads confidently and is partly wrong. The tidier the output, the harder the mistake is to see. This tutorial teaches the habit that keeps Bots safe to rely on: asking for evidence, checking in proportion to the risk, and writing the check into the brief so it happens every time. You will learn how to ask a Bot to prove its work, how to spot-check without reading everything, how to match the effort to the stakes, which errors Bots make most, how to write stop rules, and a short list of guardrails and good habits to keep. By the end you will have a checking routine that takes minutes and catches the mistakes that matter.

Difficulty: ★★★☆☆ (some practice with single Bots helps)
Tools needed: Grok Bot
Last updated: November 2026

Overview

When a colleague hands you a report, you do not usually re-do it from scratch. You skim for things that look wrong, you check the figures that matter, and you ask questions about anything odd. Over time you learn which colleagues are careful and which need watching, and you adjust.

A Bot needs the same treatment, with one important difference. A colleague who is unsure usually looks unsure. A Bot's tone is the same whether it is right or wrong. A fabricated figure and a correct one are written in the same confident voice. So you cannot judge a Bot's work by how it sounds. You have to judge it by evidence.

The Grok Bot material describes the idea. Engineers who run many Bots, such as Lauren Tan, ask a Bot to show evidence that the work is right: what was wrong before, and what is different after. You can do the same without any code. The point is not to distrust the Bot. It is to make its work checkable, so that trust can grow on something solid.

This tutorial has seven steps. You learn what kinds of error to expect. You ask for proof. You spot-check, choosing items so the check means something. You match the effort to the risk. You write the check into the brief. You add stop rules. And you finish with a short list of guardrails that make everything else easier.

Who This Is Useful For

  • Anyone using a Bot for work that other people will read or rely on.
  • People who have been fooled once by a confident, tidy, wrong answer.
  • Team owners who need a simple review standard for everyone's Bots.
  • What You Will Learn

  • The kinds of error Bots make most, and how to spot each.
  • How to ask for evidence that the work is right.
  • How to spot-check three items so the check is meaningful.
  • How to match checking effort to risk.
  • How to write the check into a brief, and how to write stop rules.
  • A list of guardrails and good habits to keep.
  • What You Need

    1. Grok Bot, with at least one Bot you use for real work.
    2. A recent piece of Bot output you can use as practice: a summary, a report, a draft.
    3. About thirty minutes for the first full check.

    Grok Bot

    The AI teammate from SpaceXAI. The official page has downloads and current plans.

    x.ai

    Information as of 14 November 2026. The idea of asking a Bot for evidence comes from how engineers who run many Bots, such as Lauren Tan, describe their work. This tutorial adapts it for people without code. Her practices are her own, and the exact tools and screens differ by version, so follow the principle.

    Step 1 — Know the errors to expect

    You check better when you know what you are looking for. Bots make a handful of errors again and again.

    ErrorWhat it looks likeWhere it hides
    Invented detailA figure, quote or source that does not existConfident, specific, unsourced claims
    Wrong numberA digit dropped, a total off, a date in the wrong formatTables, summaries, anything from a document
    Missing itemAn important thing left out entirelyLists, highlights, "all of the" claims
    Stale factA true thing that is no longer true"Recent" news, prices, who holds a role
    Wrong readingA source quoted correctly but taken to say something elseSummaries of long documents
    DriftThe Bot does something you did not ask forLong tasks and threads
    Silent skipA step not done, with "all done" reportedRoutines, bulk jobs

    The most dangerous ones are the ones that give no sign. A missing item and a silent skip look exactly like a clean result. That is why Step 2 asks the Bot to report its own gaps.


    Step 2 — Ask for the proof

    The cheapest way to check a Bot's work is to make it show how it knows. Ask for proof: links, files, before-and-after, the numbers it used.

    Add a line like this to any brief:

    Finish with an evidence list: for every figure or claim in your answer, give the source (a link, or a file name and where in it). If you could not find a source, write "no source". Then list anything you skipped, could not open or are unsure about.

    The evidence list does four things.

    1. It turns confident prose into something checkable. You can open the source.
    2. It makes inventing harder. A Bot asked to name a source for everything has to either find one or say "no source".
    3. It shows the gaps. The "skipped or unsure" list is the closest a Bot gets to a confession.
    4. It speeds up your check. You no longer search for where a figure came from.

    For work that changes things, such as tidying a folder or editing a document, ask for before and after: what it was, and what it is now. "List every file you moved, from where to where." If something is wrong, you can undo it.

    Step 3 — Spot-check, and choose well

    You do not need to check everything. You need to check the right few things. The rule is open three items yourself, and not the first three.

    The first three items are the ones a Bot is most likely to have got right, because early work is fresh and simple. Pick differently.

    1. One from the middle. Somewhere you would not look first.
    2. One that looks surprising. A number that is unusually high or low, a name you do not recognise.
    3. One that matters most. The item with the highest stakes if it is wrong.

    For each, go to the source and compare. Check the figure, the name, the date, and that the source really says what the summary claims.

    Then ask one more question: what is not here that should be? Scan the original material yourself for something important that is absent. A missing item produces no error message. This check, more than any other, is what catches the quiet failures.

    If one of your three is wrong, do not just fix that one. Look for the pattern. Is it the same kind of error? Then add a rule, and check three more.

    Step 4 — Match the effort to the risk

    Checking everything equally is slow, and checking nothing is dangerous. Match the effort to what happens if it is wrong.

    RiskExampleHow to check
    LowA rough draft for yourself, a brainstorm, a first summaryA glance, one spot-check
    MediumA report for your team, a research briefEvidence list, three spot-checks, scan for what is missing
    HighAnything with money, a public post, a message to a customer, legal or health contentA person reads every line, and every figure is verified

    A simple rule: low risk gets a quick check, high risk gets a person. If you cannot tell which it is, treat it as high.

    Two things push a job up a level without you noticing.

  • Other people will rely on it. Something harmless for you becomes risky when a colleague acts on it.
  • It cannot be undone. Sent, published, paid and deleted are all one-way. Anything one-way is at least medium.
  • Write the level on the job card for each Bot. Then the question "how much should I check?" is already answered.

    Step 5 — Write the check into the brief

    A check you have to remember is a check you will skip on a busy day. A check written into the brief happens every time.

    Add three lines to your standard brief.

    Evidence: finish by listing the sources you used, one per claim, or "no source".
    Gaps: list what you skipped, could not open or are unsure about.
    Self-check: before you answer, re-read my request and tell me in one line how you met each part.

    The third line, a self-check, is a small trick that catches drift. When a Bot has to say how it met each part of the request, parts it skipped become visible, to it and to you.

    For routines, add one more line: "If the data looks different from usual, such as a missing column or a very different total, stop and tell me instead of continuing." A routine that runs on its own should be able to say "something looks off".


    Step 6 — Stop rules

    A Bot that is going in circles, or heading somewhere it should not, should stop. A stop rule says when.

    Stop and tell me if: you have tried the same step three times; a source cannot be opened; the numbers do not add up; anything involves money, a public post or a person's private data; or you are about to do something not in this brief.

    Stop rules matter most for long tasks, routines, and threads with several Bots. They convert a runaway process into a pause, which costs a few minutes instead of a mess.

    When a stop rule fires, treat it as information, not failure. Ask what the Bot was stuck on. Often the brief was unclear, the source was unreachable, or the task was bigger than you thought. Fix that, and run it again.

    If you find yourself writing the same stop rule for every Bot, move it to your standard brief.

    Step 7 — Guardrails and good habits

    A few habits make everything above easier. They are the small rules that stop most problems before they begin.

    1. Keep "ask me first" on for money and anything hard to undo, until you trust the Bot, and for a long time after.
    2. Describe what "done" looks like, not just the first step. A Bot with a clear finish line is easier to check.
    3. Start with one tool and add more as you trust it. Each connected app widens what can go wrong.
    4. Correct in plain language. The memory sticks, so say what was wrong and what you want instead. Then check the note was learned, using the memory lesson's method.
    5. Name Bots by role and reuse them. A Bot with a history and a job card is easier to judge than a new one every time.
    6. Never share passwords or card numbers in a brief. If a Bot asks for one, stop and use a proper sign-in.
    7. Keep a one-line log. Date, Bot, job, what you found. A month later it tells you where checking pays off.

    If something goes wrong in spite of all this, the response is the same as for a routine: pause it, find the cause, fix the brief or the permission, and test again.

    Worked Example: auditing a competitor brief

    Mai asks her Research Bot for a one-page brief on a competitor, ahead of a client call. The brief arrives tidy, with five headings and a confident tone. It says the competitor "raised a funding round last quarter" and "serves around 2,000 customers".

    She had added the evidence line, so the brief ends with a list of sources. She picks three items to check, not the first three. She chooses the funding claim, which matters most, the customer figure, which looks surprising, and one product detail from the middle.

    The funding claim has a source, a news article. She opens it. The article is about a different company with a similar name. The customer figure shows "no source". The product detail checks out.

    She does not just fix the two items. She asks the Bot to redo the brief, with a new rule: "Confirm that every source is about this exact company, and put the company's website in the evidence list." She also asks for the "unsure" list, which now mentions that two sources conflict about the company's size.

    The second brief is shorter, less impressive and correct. She reads the "unsure" line aloud to the client as a question to ask. The check took ten minutes, and it kept an invented funding round out of a client call.

    Common Mistakes to Avoid

    Three habits make checking weaker than it looks.

    Mistake #1: Trusting the tone. Confident, tidy output feels reliable, and Bots sound the same when they are wrong. Fix: ask for evidence on every claim that matters, open sources yourself, and treat "no source" as a signal to verify, not a detail to skip.

    Mistake #2: Checking the first three items, or only what is there. Early items are the easiest, and a missing item gives no sign. Fix: choose one from the middle, one that looks surprising and one that matters most, and then scan the original for what should be there and is not.

    Mistake #3: Keeping the check in your head. A check you must remember is skipped on a busy day, and the same Bot gets checked differently each time. Fix: write the evidence, gaps and self-check lines into the standard brief, add stop rules, and match the effort to the risk on each job card.

    Going Further

    Make a personal error list. After a month of one-line logs, count which errors your Bots make most. Wrong numbers? Missing items? Add a targeted rule for the top two.

    Have a second Bot check the first. A Bot asked to find problems in another Bot's output, with the sources in hand, often catches things. Treat its findings as leads to verify, not verdicts.

    Review the checks themselves. Once a quarter, ask whether your checks still match the risk. Jobs creep up in stakes, and checks that were right for a draft are not right for a client deliverable.

    What to learn next. The next lesson covers trusting one skill at a time and watching usage, so that the checking you do now allows a Bot to take on more over time.

    Key Takeaways

    In this tutorial you learned:

  • A Bot sounds the same when it is right and when it is wrong. So judge its work by evidence, not by tone.
  • Know the usual errors. Invented detail, wrong numbers, missing items, stale facts, wrong readings, drift and silent skips.
  • Ask for the proof. A source for every claim, or "no source", plus a list of gaps and before-and-after for changes.
  • Open three items yourself, and not the first three. One from the middle, one surprising and one that matters most, then scan for what is missing.
  • Match effort to risk. Low gets a quick check, high gets a person, and when unsure treat it as high.
  • Write the check into the brief. Evidence, gaps and a self-check, so it happens every time.
  • Use stop rules. They turn a runaway process into a pause.
  • Keep the guardrails. Ask me first for money and anything hard to undo, describe "done", one tool at a time, correct in plain language, and never share passwords or card numbers.
  • Before you move on, a challenge. Take a recent piece of Bot output and audit it: ask for the evidence list, open three items, choosing them with care, and scan the original for something missing. Write one rule from what you find, and add the three standard lines to your brief.

    Learn AI, after work

    Track your progress, earn XP, and unlock more free tutorials in the AfterWork Academy app.

    Open this tutorial in the app

    More AI tutorials