Video version: about 20 min
Opening
About 4 minutesA Word from Mr. ZhengFictional example; numbers are illustrative
Mr. Zheng said: Give me three minutes. Explain it clearly, and I'll see if I can get two other departments to use it.
Lin Yue is no longer focused on how to do it, but on how to ensure others can take it over and use it with confidence.
In-class grading of the previous assignment: Understand the criteria first.
Three-Minute Report and Pre-emptive Questioning
About 9 minutesThree-minute presentation: Three sections
Exercise: Write your own three-minute script
Suggested time Read it aloud for about 3 minutesYou are a senior assistant who helps people prepare work presentations. Please turn my materials into a three-minute oral presentation script, approximately 300 words, divided into three sections: Problem, Methods and Data, and Request. Requirements: Use entirely conversational language; no sentence should exceed 40 characters; it should be ready to read aloud; use only the numbers I provided, without adding any effects I did not mention; clearly state 'where the human checks'; specify the sample size and which parts are backtests versus live runs; if the sample size is small, state 'conclusion is tentative'; the final sentence should ask for the other party's opinion. My materials: Task name: [Name] Original situation: [Time per instance, weekly frequency, missed items when reviewing the same checklist] Methods and division of labor: [A few sentences] Trial run data: [Sample size, number of live runs and backtests, before-and-after comparison] What colleagues said: [One direct quote] Cost: [Subscription fee, one-time investment in hours] Request: [Resources, scope, timeline, acceptance method] Presentation audience: [e.g., Department Manager]
Lin Yue's illustrative three-minute script (excerpt)Fictional example; numbers are illustrative
First, state the problem: Meeting minutes are manually created each time, taking about 90 minutes; 15 items were missed in 7 past reports.
Next, state the methods: The assistant organizes the content, and I verify responsible persons and deadlines; all external communications are personally reviewed by me.
Trial run results: 11 samples, 4 of which were live runs; combined time is about 28 minutes, missed items are about 0.4, conclusion is tentative.
My request: Try it out in the meetings of the other two departments for two weeks, with administrative colleagues verifying the content before sending it out.
Five potential follow-up questions from a superior
Suggested time Approx. 4 minutesYou are a cautious department manager who has just listened to the following three-minute report. Please ask five of the sharpest questions, one for each category: 1. Safety and data; 2. Accountability for errors; 3. Replicability by others; 4. Cost and ROI; 5. Impact on roles. Only ask questions; do not answer them for me, and do not offer suggestions. Report draft: [Paste your report draft here]
Here are my answers to the five questions. Please continue role-playing as that manager, pointing out where each answer lacks evidence, is vague, or is contradictory. Do not rewrite them for me; just list the loopholes. [Paste the questions and your answers]
Lin Yue's five answers (illustrative)
Getting Colleagues to Pilot
About 9 minutesInviting colleagues to trial: Ask first, do not broadcast
Transparent disclosure: Can the builder see your conversation with the assistant?This varies by application; refer to your current version.
Version where the builder can see
I can see the conversation between you and the assistant. This is only used to revise the script and manual, and will not be shared with others. If you do not want something recorded, tell me, and I will delete it.
Version where the builder cannot see
I cannot see the conversation between you and the assistant. I only review the records you submit to me and your answers to the five questions. You are not required to submit any part you do not wish to share.
Guided practice (15 minutes): You observe only, do not intervene.
Feedback (Five Questions): Record exact words
How to respond when a colleague resists
Fear of being replaced
Use the saved time to follow up on post-meeting action items; it only handles organization, while decisions and final approval remain with you.
Distrust of AI
You can randomly select three samples and count them against the checklist. If the numbers don't match, trust your count.
Too troublesome, no time
It only takes twenty minutes, just once. Follow the manual, and let me know where you get stuck.
Fear of being blamed or data leaks
You are the one sending it; the trial is only in drafts; use de-identified records and do not touch their actual data.
No one is willing to use it, which is also a finding
Operation Manual and Governance
About 8 minutesOne-page operation manual: Five sections
Four governance aspects: Permissions, Data, Responsibility, Audit Trail
When issues arise: Suspend first, report next, review last
Project Review and Next Project
About 6 minutesFive questions for project review
Next project: Return to the C05 terminologyConsistent with C05 guidelines
Score each of the five criteria from 1 to 5
Frequency, time consumption, clarity of rules, cost of errors, and accessibility of materials; higher scores indicate higher priority for execution.
Direction of error cost
The lighter the consequences of making a mistake, the higher the score; items involving money, contracts, clients, safety, or irreversible outcomes receive lower scores. Items scoring only 1 or 2 should be deferred.
Graduation Presentation and Class-wide Feedback
About 14 minutesFinal presentation: one card per person
Rotation method: one minute per person
Scoring sheet: maximum 22 points
Generate a presentation card from the final report
Suggested time Approx. 10 minutesPlease compress the following result report into a single presentation card, filling in the specified sections. Use only the numbers and exact wording from the report. Keep the 'measured' or 'estimated' tag for each number. For sections not covered in the report, write 'To be supplemented'; do not invent content. Quotes from colleagues must be cited verbatim. Sections: Project name; Baseline to result (three figures); Approach and human-AI division of labor (one sentence); One verbatim quote from a colleague; Unexpected findings; Areas for improvement and next steps; Automation level; Data compliance self-check. My one-page report: [Paste]
Class review: Focus on three pointsTeaching pre-set case, fictional
Focus on three points
- Can each number be traced to a specific row in the records?
- Is the colleague's quote truly verbatim? Are the unexpected findings genuinely surprising?
- Is the human oversight step clearly explained?
Common gaps
- Data section contains only positive results
- Live runs and backtests are not separated
- Colleague's quotes are rewritten as summaries
Closing, Quiz, and Assignment
About 10 minutesThree quick questions to check yourself
Question 1: What should you do when a colleague tries it out?
Question 2: When selecting the next project, how do you score the 'Cost of Error' criterion?
Question 3: What is the sequence of actions when an incident occurs?
Assignment and Project Card Updates
The ten lessons share a single thread.
Identify pain points, build processes, measure results, and bring them back to your company.
- The Dragon Warrior is merely a title; what truly matters is the boss's judgment: setting rules, defining standards, and guarding the final checkpoint.
- Tools will change, applications will update, but this methodology remains timeless.
- Next step: Return to C05 and begin your next project.
What's next
You have finished the Core lessons track
- See how other companies put AI to work: one plan per company, from public facts to the first 90 days.
- Want to turn one of your own tasks into a tool, or stuck while learning? Message John on WeChat.
- Want live sessions, a peer community and homework feedback? See the membership (the courses stay free).
Show the full transcript (video narration)
Bringing It Back to the Company
Welcome to Lesson 10, the final lesson of the AI Implementation Course: Bringing It Back to the Company. Previously, you used data to prove this workflow works. This lesson addresses how to ensure others can take it over, use it with confidence, know who is responsible if things go wrong, and how to trace actions afterward.
In this lesson, we do four things
This lesson covers four tasks. First, Persuasion: Report to your manager in three minutes and anticipate their follow-up questions. Second, Verification: Have a colleague try it out in person and record their exact words. Third, Standardization: Write an operation manual clarifying permissions, data, responsibilities, and audit trails. Fourth, Summary: Conduct a project review and select the next project based on the C05 criteria. Doing well alone is just individual skill. Bringing it back to the company means making it so others can pick it up and use it immediately, with clear responsibility for issues and full traceability afterward.
A Word from Mr. Zheng
Zheng Zong's remark was the starting point of this lesson. He said: 'Give me three minutes, explain it clearly, and I'll see if I can get the other two departments to use it.' Lin Yue's challenge was no longer about how to do the work, but how to ensure others could adopt it and use it with confidence. This case is fictional, and the numbers are illustrative; please replace them with your own data.
In-class grading of the previous assignment: Understand the criteria first.
Start with one thing: understand the criteria. The instructor selects two anonymous result reports, and the whole class grades them together using the scoring sheet. Focus on three points: Does each number indicate whether it is measured or estimated, and can you point to the specific line in the record? Are there direct quotes from colleagues and unexpected findings? Are live runs and backtests clearly separated? Let everyone understand the criteria first, then proceed with your own presentation.
Three-minute presentation: Three sections
Three-minute presentation: Only three sections. Problem (about 40 seconds): How labor-intensive was the original work? Use one number to make it clear. Methods and Data (about 100 seconds): How did you do it? How did humans and Dragon Warrior divide the work? Before-and-after comparison, where the quality gate is, and the sample size. Request (about 40 seconds): What resources do you need? What is the scope and timeline? How will it be accepted? Leave the last 20 seconds to ask: 'Do you think this scope is acceptable?' Three key principles: Few but accurate numbers (no more than five core numbers); Do not explain tool principles; State conclusions first, then the process.
Exercise: Write your own three-minute script
Now write your own three-minute script. You have already drafted it before class; here, use your phone to time yourself reading it aloud. Use this prompt to have it turn your materials into an approximately 300-word oral script, divided into three sections. Requirements include: entirely conversational language, no sentence exceeding 40 characters; use only the numbers you provided; clearly state where the human checks; specify the sample size, and which parts are backtests versus live runs; if the sample size is small, state 'conclusion is tentative.' Do not include client names or unpublished data in the script. Reading it aloud should take between two and a half and three and a half minutes.
Lin Yue's illustrative three-minute script (excerpt)
Look at Lin Yue's draft. First, the problem: meeting minutes are currently done manually, taking about 90 minutes. In the last seven issues, 15 items were missed. Second, the solution: the assistant organizes the content, Lin Yue verifies the responsible persons and deadlines, and she personally approves all external communications. Trial results: 11 samples, 4 of which were fully executed, averaging about 28 minutes with approximately 0.4 missed items. Conclusion is tentative. Finally, the request: trial this in the weekly meetings of two other departments for two weeks. Have administrative colleagues verify the output before sending. Review using the same metrics after two weeks, and ask Mr. Zheng to confirm the scope.
Five potential follow-up questions from a superior
The biggest fear in a report is being stumped. First, have it role-play as a cautious department manager to ask five of the sharpest questions based on your report, one per category: safety and data, accountability for errors, replicability by others, cost and ROI, and impact on roles. The prompt requires it to only ask questions, not answer them for you. Write out your answers to the five questions, then send them to it. Have it continue role-playing as the manager to point out where there is no evidence, where things are vague, and where there are contradictions. It only lists loopholes; it does not rewrite them for you.
Lin Yue's five answers (illustrative)
Lin Yue's answers to the five follow-up questions are illustrative. Is it safe: Records are desensitized first, excluding client names and amounts, with access logs specifying read-only access to one folder and draft creation only. Who is responsible for errors: The sender is responsible, which is Lin Yue herself, checking against a seven-point list before every send. Can others replicate this: There is a one-page operation manual; ask a colleague who hasn't done it before to try it. Cost and ROI: Figures for subscription and one-time investment; you fill in your own numbers, calculate the payback period in weeks using the C09 formula, and specify whether it is the monetary or hourly version. Will it replace people: Saved time is used to follow up on post-meeting action items; role arrangements are not decided by the assistant; the final approval gate remains human. Each answer uses only data from records or policy documents. If you don't know, write 'I will verify.'
Inviting colleagues to trial: Ask first, do not broadcast
Testing it yourself and having others use it are very different. When inviting a colleague for a real-person trial, ask them first instead of mass-inviting them. Clearly state in a few sentences: what it does, how long it takes, who it won't affect, how data is protected, and who can see your conversation. It uses anonymized records and does not involve your colleagues' personal materials. If they decline, simply say, 'Okay, sorry to bother you.' Do not try to persuade them.
Transparent disclosure: Can the builder see your conversation with the assistant?
The question colleagues most want to ask, but feel hesitant to, is whether you can see their conversation with the assistant. Address this proactively during the invitation rather than waiting for them to ask. First, verify whether the builder of your dedicated assistant can see others' conversations in the application you are using, as this varies by application; refer to your current version. Then, state the facts clearly, choosing one of the two versions. A reminder: deletions only remove content visible to you. Whether the service provider retains data, and for how long, depends on company policies and application settings; do not make promises on behalf of the service provider. If you are unsure, consult the IT department first, then inform your colleague.
Guided practice (15 minutes): You observe only, do not intervene.
Guided practice (15 minutes). In the first minute, say only: 'Please follow this page of the manual and do it yourself.' From minutes 2 to 14, watch quietly, use a timer to record the total time, and note only three types of moments: where they got stuck, where they smiled, and where they frowned. Record the timestamp and their exact words for each. If they ask you a question, reply only with: 'What do you think should be done?' and record what they asked. At minute 15, say: 'Okay, thank you. I have five questions for you.' Throughout the process, do not provide hints, explanations, or click things for them.
Feedback (Five Questions): Record exact words
After the trial, ask five questions. 1. What did you use it for? 2. At which moment did you feel it understood your work, and which sentence or step was it? 3. At which moment did you feel it was talking nonsense or answering irrelevantly? 4. Will you use it in the future? 5. Describe it in one word. There is only one recording rule: record exact words. Do not summarize for them or sugarcoat. If they say 'It was okay,' write 'It was okay.' Do not answer for them, do not rush to explain, and do not persuade them that it is actually useful. Exact words are more valuable than your summary.
How to respond when a colleague resists
Colleagues have concerns. How do you respond? Fear of being replaced: The time saved by this task is used for following up on post-meeting action items; it only handles organization, while decisions and final approval remain with you. Distrust of AI: You don't need to trust me, nor do you need to trust it. You can randomly select three samples and count them against the checklist. If the numbers don't match, trust your count. Too troublesome or no time: It only takes twenty minutes, just once. Follow the manual, and let me know where you get stuck. Fear of errors, blame, or data leaks: The sender is you, not the AI. The trial is only in drafts. We use de-identified records and do not touch your actual data. Responses must be specific: clarify what it does, what it doesn't do, and who is responsible for approval.
No one is willing to use it, which is also a finding
No one is willing to use it, which is also a finding. Write down the reason, which type is it: didn't dare to send it, sent it but no one replied, or used it but found it useless. The next steps differ for each. If you didn't dare to send it, practice once with the colleague you know best. If you sent it but no one replied, try asking in person at a different time. If you used it but found it useless, go back to the third question of the Five Questions, and see which part made him feel it was useless. Writing that no one is willing to use it is not shameful. Fabricated positive reviews are what will undermine your subsequent report.
One-page operation manual: Five sections
One-page manual, five sections. Preparation: Materials, de-identification, permitted tools; de-identification must include an example. Steps: Submit to the dedicated assistant, wait for the first draft, check against the acceptance checklist item by item, handle missing items per common issues, and send only after confirming everything is correct. Mandatory review: Responsible person, deadline, amount, and content sent externally. What to do if issues arise: Who to contact, how to record, whether to suspend usage. Records: Each entry: Date, Time Spent, Missed Items, Notes. Validation standard: Have a colleague who hasn't done this before perform the task independently, without asking you, and achieve a qualified result.
Four governance aspects: Permissions, Data, Responsibility, Audit Trail
Four governance aspects: Permissions, Data, Responsibility, Audit Trail. Permissions: List names, not 'everyone'; specify which files can be read, whether sending is allowed, and whether file modification is allowed; default to least privilege, start with 'draft only, no sending'. Data: Specify what can and cannot be submitted; mandatory data must be de-identified first; do not paste account numbers or passwords. Responsibility: Each process has one responsible person, name written in the manual; responsibility for sent content lies with the sender, not the tool. Audit Trail: One record line per use; version numbers for script changes. Data classification, retention periods, and external notification requirements follow company policies and information security/legal department regulations; do not decide independently.
When issues arise: Suspend first, report next, review last
When issues arise, the sequence is fixed: Suspend first, report next, review last. For example, if a draft is sent by mistake or unauthorized data enters the tool, stop the spread immediately. Report next, using three sentences to clarify: What happened (time, process, phenomenon); what data is involved (approximate count, whether it includes red-line data); what actions have been taken (what has been suspended, who has been notified, who is awaiting a response). Analyze causes after the incident is resolved. Before the pilot, also submit a report: Who approved, scope, duration, data category, and approval method (written or recorded oral).
Five questions for project review
Project review using five questions: What was the biggest surprise this time? Which assumption was wrong? Which step took the most time? What would you fix first next time? What would you advise colleagues to avoid? Conduct this review in the week the project ends, correlating with your unexpected findings. Lin Yue's review example: Colleagues cared most about the 'to-be-confirmed' column, not saving time; she thought de-identification was simple, but colleagues spent three minutes stuck on it; the most time-consuming step was organizing case files for the first time; next time, write the de-identification example in the manual first.
Next project: Return to the C05 terminology
For the next project, do not choose based on intuition. Return to C05: the guidelines remain exactly the same—score frequency, time consumption, clarity of rules, cost of errors, and accessibility of materials from 1 to 5, with higher scores indicating higher priority. Note the direction of error cost: the lighter the consequences of a mistake, the higher the score; items involving money, contracts, clients, safety, or irreversible outcomes receive lower scores. Tasks with an error cost score of only 1 or 2, regardless of the total score, should be deferred until safeguard measures are in place. Xiao-Tang's next step is to expand the project to the technical support team of another product line.
Final presentation: one card per person
Final presentation: one card per person. The card is structured from top to bottom: project name; 1. Baseline to results, three numbers indicating actual measurements or estimates; 2. Method and human-machine division of labor, one sentence; 3. A direct quote from a colleague; 4. Unexpected findings; 5. Areas for improvement and next steps, specifying the exact location. The final two sections are: automation level, and data red-line self-check (passed or failed). During the presentation, read directly from the card without adding any extra content.
Rotation method: one minute per person
Rotation method. Each group of five to six people selects a timekeeper, rotating the role. Each person speaks for one minute, reading directly from the card without adding extra content. Listeners perform two tasks: write one specific comment of their own, such as 'I will borrow your step X,' avoiding vague praise like 'good job'; then draw a number and ask the presenter to point to the corresponding line in their notes. A six-person group takes approximately nine minutes. Absent members submit their presentation card asynchronously along with a one-minute audio recording, which must also include one specific piece of feedback from others.
Scoring sheet: maximum 22 points
The scoring rubric has 11 items, each worth 0–2 points, for a maximum of 22. We focus on four key items here. Clear problem statement: baseline figures are provided, and each is marked as measured or estimated. Credible data: figures come from records, sample sizes are stated, and live runs are separated from backtests. Honesty: shortcomings, causes, and unexpected findings are reported as they occurred. Verifiable data: pick any two figures, and you can trace them to specific rows in the records. Scores of 18–22 are excellent, 13–17 are passing, and 12 or below suggest rework before presentation. The score highlights gaps; it does not rank you.
Generate a presentation card from the final report
Complete the presentation card before class. Use this prompt to generate the card from your final result report, then revise it yourself. It must use only the numbers and exact wording from the report, keeping the 'measured' or 'estimated' tag for each figure. For sections not in the report, write 'To be supplemented'; do not invent content. Quotes from colleagues must be cited verbatim. After generation, verify each section yourself.
Class review: Focus on three points
The class reviews two typical cases. Representatives from each group come forward. The instructor selects one strong example and one with clear gaps. The class scores them live using the rubric. This is not about finding fault, but about understanding which aspects determine credibility for the same project. The review focuses on three points: Can each number be traced to a specific row in the records? Is the colleague's quote truly verbatim? Are the unexpected findings genuinely surprising? Is the human oversight step clearly explained? There are three common gaps: Data sections contain only positive results; live runs and backtests are not separated; colleague quotes are rewritten as summaries.
Quiz Question 1
Question 1: What should you do when a colleague tries it out? A. Observe and explain. B. Do not intervene or teach; only note moments of confusion, smiles, frowns, and record their exact words. C. Let them finish reading your report first. The answer is B. Self-testing is not the same as real-user testing. Observing is more honest than asking afterward.
Quiz Question 2
Question 2: When selecting the next project, how do you score the 'Cost of Error' criterion? A. The higher the cost, the higher the score. B. The lighter the consequences of making a mistake, the higher the score; projects scoring only 1 or 2 points should be set aside for now. C. This criterion does not need to be scored. The correct answer is B. This is consistent with C05 guidelines: the higher the score, the more suitable the project is for prioritization.
Quiz Question 3
Question 3: What is the sequence of actions when an incident occurs? A. Conduct a post-mortem first, then report. B. Suspend operations first, then report, followed by a post-mortem. C. Identify the cause first, then decide whether to report. The correct answer is B. Stop the escalation first, then inform those who can handle it; address the cause afterward.
Assignment and Project Card Updates
Submit these assignments: A timed three-minute presentation script, along with answers to five follow-up questions. A colleague's trial record, specifying the verbatim answers to the five questions, three key moments, and how you ensured transparent communication. A one-page operation manual, tested by a colleague who has not performed the task before. A governance checklist and pilot registration. Five post-mortem questions for the project, and a scoring sheet for the next project. A presentation card, with self-assessment based on the scoring rubric. Submit the first four items and the presentation card within one week after the course; submit the post-mortem and scoring sheet during the week the pilot concludes. The AI can help polish your work, but it cannot fabricate your timing records or your colleagues' verbatim responses.
The ten lessons share a single thread.
To wrap up: over these ten lessons, there’s really just one thread: identify the pain point, build the workflow, measure the results, and then bring it back to your company. You pick the most time-consuming task from your candidate list, record the baseline, break down the process, and decide what to delegate to the Dragon Warrior versus what you must handle yourself. You create a custom assistant, get it running, and then use data and colleague feedback to justify bringing it back to your organization. The Dragon Warrior is just a title; what really matters is the boss’s judgment: setting rules, defining standards, and guarding the final gate. Tools change, applications update, but this method remains timeless. Your next steps are threefold: the ten-lesson course (8–10 minutes per lesson), the mobile-friendly version for those who only use phones, and most importantly, return to C05 to start your next project. Thank you for watching. This video’s narration is AI-generated.