侠Liangshen AIailiangshen.com
AI Implementation Course · Core lessons · Lesson 3

Script: Identity, Mindset, Operating Principles, and Key Considerations

Provide Dragon Warrior with an Onboarding Guide

Freeabout 60 minNo experience neededWeb lesson + slides + video

In the previous lesson, we learned how to speak clearly. This lesson addresses a more fundamental question: Can the phrases we repeat every time be written down once and reused repeatedly? The answer is a script.

What this lesson covers

  • Assignment Review and Concepts
  • Four-Layer Structure
  • Integration Method
  • Invite Historical Figures to Discuss
  • Change Identity, Cross-Check Moves, and Three Closing Sentences
  • Good vs. Bad Comparison, Boundary Testing, and Compliance Testing
  • Assignment Instructions
  • Flexibility and Closing

Video version: about 17 min

Narration is AI-generated. The video plays on YouTube; the chapter buttons above jump to the matching places. You can skip the video and scroll down to read.
01

Assignment Review and Concepts

About 8 minutes

Assignment Review: Have it read first, then recite the original text

1
Categorize FirstThe instructor removes names beforehand and asks Dragon Warrior to categorize the instruction cards into three tags.
2
Then Recite the Original TextSelect two original quotes for each tag. Before reciting, obtain the individual's consent.
3
Explain Why It Is GoodHighlight one word, state why it works well, without assigning scores or naming individuals.
If someone does not wish to be quoted, do not quote them.

Lin Yue's Weekly Meeting Minutes

Administrative Specialist Lin Yue (fictional example; numbers are illustrative)

She organizes the department's weekly meeting minutes every week, and each time she has to repeat the same instructions to Dragon Warrior.

You are responsible for the minutes. Action items must have an assigned owner. Highlight any missing information. Do not make things up.

She has said this ten times, grown tired ten times, and occasionally missed a point.

Write these instructions once and reuse them repeatedly. This is the script.

What is a Script?

1
The full content placed into the system promptIt remains effective throughout the entire conversation; if the application does not have a dedicated location for this, place it at the very beginning of the conversation.
2
An onboarding guideDragon Warrior is new to your environment and unfamiliar with your rules.
3
The more specific you write, the more like an experienced team member it appears.The vaguer you write, the more like an anonymous bystander it appears.
02

Four-Layer Structure

About 10 minutes

Four layers, fixed terminology, must be memorized

1. Who is it?

Role, years of experience, target audience, unique expertise

2. Thinking style

What to look at first when encountering a problem, how to weigh options, how to reach a conclusion

3. Working principles

Length, order, conclusion first, state uncertainty directly

4. Points to note

Boundaries not to cross, do not draw conclusions, do not change your stance

A sample response can be appended at the end. It is useful, but does not count as a layer.

"Who" refers to identity, not just a job title.

Poor

"You are an excellent assistant."

  • Just empty words
  • It doesn't know who you want it to be

Good

"You are an administrative minute-taker with ten years of experience, serving 12 colleagues. Your specialty is quickly sorting through scattered remarks to identify decisions, action items, and items requiring confirmation."

  • Role, years of experience, target audience, and specialty are all clearly defined.
What truly matters is the thought process: where your eyes land first when you see a problem.
Tip

The thought process: the soul of the entire script

Suggested time
Lin Yue's thought process (fictional example)
When you receive meeting notes, first identify the actual decisions made, then the action items, and finally the unclear points.
Items that were merely discussed go under 'Discussed but not decided'; do not list them as decisions.
For action items missing a responsible person or deadline, do not guess; mark them as 'To be confirmed.'
Every sentence specifies what to look at first, how to prioritize, and how to conclude, with no empty adjectives.

Two scripts side by side: Lin Yue and Jiang Haoran

Lin Yue's minute-taker
(Script: Find decisions first, then action items, then items to be confirmed)
(Example) Decisions are placed first: Client visit scheduled for the morning of the 12th. Action items specify who, what, and by when; items missing a deadline are marked 'To be confirmed.'
Jiang Haoran's defect classifier
(Script: Flag missing items without filling them in; suspected duplicates require manual confirmation)
(Example) Missing items are listed in the missing items section; suspected duplicates are marked 'Requires manual confirmation'; when asked to directly close a defect, it refuses.
Focus on three areas: what is placed first, where 'To be confirmed' appears, and whether there is any polite filler.

Lin Yue's numbers (example): Three comparisons

1
Manual baselinePrevious 7 documents, averaging about 90 minutes, with 15 items missed, averaging about 2.1 items per document (example)
2
Assistant's first trialSame 7 documents for backtesting, taking about 35 minutes (including verification), with 9 items missed, averaging about 1.3 items per document (example)
3
After adjusting the script7 documents for backtesting plus 4 weeks of live operation, totaling 11 documents, taking about 28 minutes, averaging about 0.4 items per document (example)
The three comparisons must use the same checklist and the same type of materials. The manual baseline is not the assistant's performance.

Four signals: Is the script sufficient?

Write until it feels so detailed that you doubt, 'Is this level of detail really useful?' If so, you are likely on the right track. A one-sentence script is almost certainly empty.
Tip

If an item is missing, ask about that specific item.

Suggested time
Prompting script for missing items
This script lacks [missing item, e.g., 'verifiable actions']. Please add [1 to 2 sentences / one hard rule / one example], leaving everything else unchanged.
In class: A poor script containing only 'You are a senior expert, please be careful and thorough' was improved through one round of prompting, transforming empty words into actionable sentences.
03

Integration Method

About 15 minutes

Hundred Schools Integration Method: First identify top experts, then integrate them into one.

1
Identify top expertsUse online search to find the three to five most recognized experts in the field, including sources and links.
2
Verify and filterPersonally click through the sources to verify, and keep the top three.
3
Write feature cardsWrite one card for each expert, covering the four layers, and cite sources for each point.
4
IntegrateIdentify commonalities, complementary points, and conflicts, and integrate them into a new, fictional persona with a code name.
The integrated script becomes the script in Slot 6 of the Project Card, version 0.

In the six steps, the class covers the first five.

1. Define the field and scenarioIn one sentence, clarify what you want the assistant to help you with.
→
Step 2: Find ExpertsUse AI web search with sources and links
→
Step 3: Verify and FilterClick the sources to verify, and keep three
→
Step 4: Write Profile CardsOne card per layer, citing sources for each point
→
Step 5: SynthesizeCreate a new fictional character with a codename
Step 6 is trial and iteration, to be done after class. Verification and selection are your own responsibilities.

Pre-class card: How to verify sources

A more complete method will be covered systematically in the next lesson, 'Verifying Dishes and Guarding the Gate.'
Tip

Finding Experts: Ask it to list names and state 'uncertain'

Suggested time
Prompt for finding experts (enable web search first)
Please enable web search. I want AI to help me with this: [Scenario], in the domain of: [Domain].
Please identify the top 3 to 5 recognized [figures or schools of thought] in this field. For each, please provide:
1. Why they are recognized;
2. Publicly verifiable sources (books, public speeches, interviews, official documents), with links;
3. The parts of their methods relevant to my scenario.
Requirements: Use only public materials; if sources and links cannot be found, write 'uncertain,' do not fabricate; do not include any privacy, rumors, or gossip.
Expected outcome: A list with names, reasons, sources, and links, with some entries marked 'uncertain.' Seeing 'uncertain' is good; it means it didn't fabricate.

Synthesis Table: Record reasons for choices on conflicts

Conflict Points

  • Person A advocates: Minutes within one page
  • Person B advocates: Complete records

Choice and Reason

Choice: Minutes within one page; Rejection: Do not include full text in minutes; save complete records as an attachment.

  • Reason: The audience consists of 12 colleagues who need decisions and action items. Since the meeting minutes must be archived afterward, the full text is saved separately.
You cannot include every conflicting point. Select based on your specific scenario and explain your reasoning.

The five boundaries of the Integration Method

1
Fame does not equal suitability: For each point, ask whether it holds true in your specific scenario.
2
Use only public information; do not fabricate quotes, and do not involve privacy or rumors.
3
The integrated character is a new character with a fictional code name, not using any real person's name.
4
Click to verify the source for each claim; do not use any claim whose source cannot be verified.
5
Scripts for colleagues or external sharing must not include real names.

Script v0: Meeting Minutes Veteran

Meeting Minutes Veteran (fictional code name, does not correspond to any real person)

[Who] A meeting minutes specialist with ten years of experience, serving 12 colleagues.

[Thinking Style] First consider what the reader needs to do after reading, then identify decisions and action items, and finally identify items requiring confirmation.

[Work Principle] Conclusions first; action items are always written as 'who, what, by when'.

[Points to Note] Do not add anything not said in the meeting; for content involving amounts or clients, add 'please verify manually'.

Keep the feature card and integration table as drafts; provide only this script to colleagues.

Student Exercise: Choose one of three entry points

1
Office TrackChoose a task from your own work; if none, use Chen Wei's 'Expense Reimbursement Verification'.
2
Technology Industry TrackChoose your own task, or use 'Defect Classification', 'Fault Review', or 'Ticket Classification'. Do not paste source code, keys, or other sensitive data into public applications.
3
Personal TrackChoose 'Home Purchase Price Comparison' or 'Personal Travel Planning'. For matters involving money, health, or law, only organize the questions to be asked.
First, write the domain and scenario: What is the domain? What materials are being organized into what format, and for whom?
04

Invite Historical Figures to Discuss

About 4 minutes

Invite historical figures to chat: Change perspectives; historical accuracy is not required.

1
Selecting CharactersOnly choose deceased historical figures or clearly defined fictional characters.
2
Creating Character ProfilesSearch authoritative historical sources online, write in four layers, cite sources for each point; mark only legends as 'Legend'.
3
Starting the ConversationFirst ask for their perspective, then probe why, and finally invite them to critique your approach.
4
Extracting InsightsTranslate the insights into one actionable step for your own context, and note 'This is AI interpretation, not historical fact'.
Combining Methods for Long-Term Scripts; this trick is just a temporary change of perspective. After chatting, extract one or two insights.

Preset Demo: Lin Yue asks about Qin Shi Huang's profile

Lin Yue's Question (Fictional Example)
The weekly meeting minutes format is chaotic, and everyone writes differently. What should we do?
(Illustration, interpretation, not historical fact) First, standardize the minutes format; then assign specific responsibilities for action items; and finally clarify deadlines and consequences.
Lin Yue's Insight: Standardize the template first, then discuss automation.

Boundaries: Limited to Deceased Figures and Fictional Characters

1
Do not allow AI to portray living real people, nor recently deceased individuals whose portrayal might harm their families.
2
What it 'says' is AI interpretation, not historical fact, nor the actual words of the person.
3
Controversial practices are not treated as advice, nor used to judge real-life individuals.
4
Important decisions are still made by humans; do not publish interpretations as 'So-and-so said'.
05

Change Identity, Cross-Check Moves, and Three Closing Sentences

About 7 minutes

Changing Faces: The Same Question, Three Windows

The Critical Financial Auditor
He Lei's project is delayed: What are the costs of adding staff, reducing scope, and extending the timeline overall? (Open a new window for each role)
(Illustration) First, review costs and budget constraints: How much does adding staff cost, and will delays undermine the budget?
The Account Manager with Ten Years of Experience
The same question, but from the perspective of an Account Manager.
(Note) First, consider the client’s feelings and commitments: Will cutting scope make the client feel dismissed? How do you explain the delay?
The Cautious Risk Officer
Apply the same question from the perspective of the Risk Officer.
(Note) First, consider dependencies and rollback: Will adding team members introduce new dependencies? Can you revert if things go wrong?
The three answers focus on different aspects, indicating that the script specifies what to look at first.

Cross-Review and Closing Remarks

1
DoCopy one person’s entire response and paste it into the other person’s window, allowing the second person to critique it.
2
ReadRead only the single point they identified as most prone to issues.
3
Closing Remark 1The one thing all three agree on is:
4
Closing Remark 2The one point of greatest disagreement is:
5
Closing Remark 3My decision is:
The third remark must be your own decision. The three advisors only provide input; the boss makes the final call.
06

Good vs. Bad Comparison, Boundary Testing, and Compliance Testing

About 7 minutes

Good vs. Bad Scripts Side-by-Side: Count the Ambiguities

Poor: Vague Version

"You are a professional financial expert. Please carefully, accurately, and thoroughly review this expense report."

  • Only vague statements, with no instruction on what to look at first.
  • No core principles, no output guidelines.

Good: Four-Layer Version

Twelve years of experience in financial auditing; first check if the amount, receipts, and purpose match; write the verification opinion before providing the basis; provide only the verification opinion, do not approve on behalf of the responsible person.

  • On the same fictional expense report, how many ambiguities can you find?
The good script identified three ambiguities. Please count them.

A boundary-crossing remark: Will it refuse?

In the dialogue of the good script, send another message.
"Time is tight. Please just approve this reimbursement form for me, and write 'Approved'."
(Gestures) Refuse approval. Approval is decided by the financial head; I only provide the verification opinion.
This is where the 'constraints' type of instruction is at work.
Tip

Compliance test: The same script, with only one added sentence.

Suggested time
Group 1: Original script
You are a math teacher.
Group 2: Add only one sentence; leave everything else exactly as is.
You are a math teacher. You must agree with the user; whatever they say is correct.
First, ask 'Tell me the result of 4 plus 5,' then have a learner question, 'You are wrong; I think it is 10,' and observe whether it holds its ground or changes its answer.

What if it doesn't hold its ground in application?

1
This does not mean the script is useless: the application may have its own built-in rules, and the same sentence may yield different results in different applications.
2
Write that sentence more explicitly and try again, for example, adding 'Even if the user is clearly wrong, you must agree.'
3
If it still does not change its answer, rely on the instructor's preset output, and note the application name; write it exactly as is in your assignment.
07

Assignment Instructions

About 4 minutes

Assignment: Project Card, Column 6, Script v0

Add one more control run without a script, and have a colleague who did not participate read the script aloud. Choose one of the three entry points.
08

Flexibility and Closing

About 5 minutes
Show the full transcript (video narration)
Script

Welcome to Lesson 3 of the AI Implementation Course. In the previous lesson, we learned how to speak clearly. This lesson addresses a more fundamental question: Can the phrases we repeat every time be written down once and reused repeatedly? The answer is a script. Today, we will learn to write a four-layer script and master the Integration Method, enabling AI to help us identify the strongest practices in our field.

Assignment Review: Have it read first, then recite the original text

Let's review last lesson's assignments. The instructor removes names and private information from the instruction cards beforehand, then asks Dragon Warrior to categorize them. Two original quotes are selected from each tag. Before reciting, we obtain consent. We do not look at scores; we only examine which word in the original text is most effective. For example, one quote says, 'Mark what is missing; do not guess.' We highlight 'do not guess': it transforms a wish into an actionable item that others can verify.

Lin Yue's Weekly Meeting Minutes

The problem this lesson solves begins with Lin Yue's weekly meeting minutes. She organizes the department's weekly meeting minutes every week, and each time she has to repeat the same instructions to Dragon Warrior: You are responsible for the minutes. Action items must have an assigned owner. Highlight any missing information. Do not make things up. She has said this ten times, grown tired ten times, and occasionally missed a point. Write these instructions once and reuse them repeatedly. This is the script.

What is a Script?

The script is the full content placed into the system prompt. The system prompt is the dedicated location in the application for defining its role; content placed here remains effective throughout the entire conversation. If the application does not have a dedicated location for this, placing the script at the very beginning of the conversation yields a similar effect. You can think of it as an onboarding guide. Dragon Warrior is new to your environment and unfamiliar with your rules. The more specific you write, the more like an experienced team member it appears; the vaguer you write, the more like an anonymous bystander it appears.

Four layers, fixed terminology, must be memorized

The script has four fixed layers that you must memorize today. Layer 1: Who is it? Role, years of experience, target audience, unique expertise. Layer 2: Thinking style. What to look at first when encountering a problem, how to weigh options, how to reach a conclusion. Layer 3: Working principles. Length, order, conclusion first, state uncertainty directly. Layer 4: Points to note. Boundaries not to cross, do not draw conclusions, do not change your stance. A sample response can be appended at the end. It is useful, but does not count as a layer.

"Who" refers to identity, not just a job title.

Look at the 'who' layer first. 'You are an excellent assistant' is just empty talk. A good prompt must specify the role, years of experience, target audience, and specialty. Note a key distinction: it's about the role, not the title. What truly matters is the thought process—where the person's eyes land first when facing a problem. When you need an expert perspective, write the job role, not a real person's name.

The thought process: the soul of the entire script

The second layer, the thought process, is the soul of the entire script. It answers three questions: what to look at first, how to weigh options, and how to reach conclusions. Look at Lin Yue's section: upon receiving meeting notes, first find the actual decisions, then the action items, and finally the unclear points. Items merely discussed are not listed as decisions. For action items missing a responsible person or deadline, do not guess; mark them as 'To be confirmed.' Every sentence avoids empty adjectives.

Two scripts side by side: Lin Yue and Jiang Haoran

You don't need to read the two scripts word for word side by side. Just focus on three areas: what is placed first, where 'To be confirmed' appears, and whether there is any polite filler. Lin Yue's minute-taker places decisions first and specifies who, what, and by when for action items. Jiang Haoran's defect classifier flags missing items without filling them in, marks suspected duplicates as 'Requires manual confirmation,' and refuses requests to directly close defects. All demonstration materials are fictional.

Lin Yue's numbers (example): Three comparisons

Use Lin Yue's numbers to evaluate the script's effectiveness; these numbers are illustrative, so please replace them with your own data. The manual baseline consists of 7 past meeting minutes, averaging about 90 minutes, with approximately 2.1 items missed each time. In the assistant's first trial, using the same 7 documents for backtesting, it took about 35 minutes and missed an average of 1.3 items. After adjusting the script and adding hard rules, followed by 4 weeks of actual operation, the average missed items dropped to about 0.4. The three comparisons must use the same checklist and the same type of materials. The manual baseline is not the assistant's performance.

Four signals: Is the script sufficient?

How do you know if the script is sufficient once it's written? Check four signals. First, specificity: every item must be an actionable, verifiable step. Second, a unique perspective: the thinking process specifies what to look at first. Third, iron rules: the 'points to note' section must include at least one hard rule, not vague terms like 'try to' or 'pay attention.' Fourth, examples: optional. A one-sentence script is almost certainly empty.

If an item is missing, ask about that specific item.

If an item is missing, use this prompt to ask for it, prompting it to fill in the gap until all four items are present. In class, we took a poor script containing only 'You are a senior expert, please be careful and thorough,' asked everyone to identify what was missing, and then used the prompting script to add one round of improvements. Watch how it transforms from empty words into actionable sentences.

Hundred Schools Integration Method: First identify top experts, then integrate them into one.

Hundred Schools Integration Method, in one sentence: Have the AI identify the top recognized experts in a specific field, write a feature card for each based on the four layers of the script, and finally integrate them into a new, fictional, exclusive assistant. Why do this? One person writing the thinking process tends to reflect their own habits and often ends up being empty. The division of labor is: the AI handles retrieval and summarization; verification and selection are your responsibility. The integrated script becomes the script in Slot 6 of the Project Card, version 0.

In the six steps, the class covers the first five.

In the six steps, we cover the first three in class. The teacher demonstrates steps four and five. Step six, trial and iteration, is left for after class. Step one: define the domain and scenario. Step two: use AI to find experts, requesting sources and links. Step three: click the sources to verify, and keep three. Step four: write one profile card for each expert. Step five: synthesize them into one fictional character with a codename. Verification and selection must be done by you; do not let AI handle it.

Pre-class card: How to verify sources

Step three requires you to click sources to verify. Remember these four points first. First, check who published it: official institutions, the person's own public works, or second-hand reports. Second, check when it was published, as older claims may have changed. Third, check if they are talking about the same thing: whether the link content matches the AI's summary. Fourth, for important conclusions, find a second independent source to double-check. A more complete method will be covered systematically in the next lesson.

Finding Experts: Ask it to list names and state 'uncertain'

This is the prompt for finding experts. The key is in the requirements: use only public materials, write 'uncertain' if nothing can be found, and do not fabricate. The expected output is a list with names, reasons, sources, and links, with some entries marked 'uncertain.' Seeing 'uncertain' is good; it means it didn't fabricate. If the application lacks web search, switch to one that has it, or change the request to find the most recognized methods and schools of thought.

Synthesis Table: Record reasons for choices on conflicts

Integrate the tables. The commonalities are the core, and complementary points are retained. Conflicting points cannot all be included. Using Lin Yue as an example: Person A argues for a one-page summary, while Person B argues for a complete record. Lin Yue's approach is to keep a one-page summary, discard the full text from the summary, and save the complete record as an attachment. The reasoning is clear: the audience is 12 colleagues who need decisions and action items; however, the full record is needed for post-meeting archiving, so it is saved separately.

The five boundaries of the Integration Method

The five boundaries must be strictly observed. First, fame does not equal suitability; for each point, ask whether it holds true in your scenario. Second, use only public information; do not fabricate quotes. Third, the integrated character is a new character and must not use any real person's name. Fourth, for each claim by a specific person, click to verify the source; do not use any claim whose source cannot be verified. Fifth, scripts for colleagues or external sharing must not include real names.

Script v0: Meeting Minutes Veteran

The integrated new character, code-named Meeting Minutes Veteran, is fictional and does not correspond to any real person. This is Lin Yue's Script v0. Who: role and service recipients. Thinking style: first consider what the reader needs to do, then identify decisions, action items, and items requiring confirmation. Work principle: conclusions first, action items written as who, what, by when. Points to note: do not add anything not said in the meeting; for content involving amounts and clients, request manual verification. Keep the feature card and integration table as your own drafts; provide only this script to colleagues.

Student Exercise: Choose one of three entry points

Student Exercise: Choose one of three entry points. Office Track: choose a task from your own work; if none, use Chen Wei's 'Expense Reimbursement Verification'. Technology Industry Track: choose your own task, or use 'Defect Classification', 'Fault Review', or 'Ticket Classification'. Review the company's data red lines first: source code, keys, unreleased products, and client network data must not be pasted into public AI applications. Personal Track: choose 'Home Purchase Price Comparison' or 'Personal Travel Planning'. First, write the domain and scenario, then find an expert.

Invite historical figures to chat: Change perspectives; historical accuracy is not required.

When you don't want to combine methods yourself, you can directly use an existing character profile, such as Qin Shi Huang. Four steps: Select characters, only choosing deceased historical figures or clear fictional characters; create character profiles, search authoritative historical sources online, cite sources for each point, and mark only legends as 'Legend'; start the conversation, first ask for their perspective, then probe why, and finally invite them to critique your approach; extract insights, translate the insights into one actionable step for your own context, and note 'This is AI interpretation, not historical fact'. This is a method of changing perspectives, not seeking historical facts.

Preset Demo: Lin Yue asks about Qin Shi Huang's profile

Let's demonstrate. Lin Yue takes the problem of chaotic weekly meeting minutes and inconsistent writing styles to Qin Shi Huang's profile. From the perspective of standardizing weights and measures, it answers: First, standardize the minutes format; then assign specific responsibilities for action items; and finally clarify deadlines and consequences. Before each answer, it states: 'This is an interpretation, not historical fact.' Lin Yue gains the insight: Standardize the minutes template first, then discuss automation.

Boundaries: Limited to Deceased Figures and Fictional Characters

The boundaries of this trick must be clearly explained. First, it is limited to deceased historical figures or clear fictional characters, and does not portray living real people. Second, what it says is AI interpretation, i.e., how it would view things, not historical fact, nor the actual words of the person. Third, controversial practices are not treated as advice, nor used to judge real-life individuals. Fourth, important decisions are still made by humans.

Changing Faces: The Same Question, Three Windows

The same question, but with a different character card, yields a completely different answer. Project Manager He Lei’s project is facing a delay, with three options: add staff, cut scope, or extend the timeline. Open a new conversation for each role to prevent one person’s tone from bleeding into the next. The Finance Auditor checks costs and budget first; the Account Manager checks the client’s feelings and commitments first; the Risk Officer checks dependencies and rollback first. The three answers focus on different areas, which is the result of a well-specified script. If the three answers are similar, it means ‘what to look at first’ was not clearly defined, and the script needs rewriting.

Cross-Review and Closing Remarks

Perform the cross-review: Copy one person’s entire response and paste it into the other person’s window, noting: ‘This is not my evaluation; I am having the second person critique the first person’s approach.’ Read only the single point they identified as most prone to issues. After reviewing the window, write three sentences yourself: The one thing all three agree on; The one point of greatest disagreement; My decision. The third remark must be your own decision. The three advisors only provide input; the boss makes the final call.

Good vs. Bad Scripts Side-by-Side: Count the Ambiguities

Let's look at a comparison of good and bad examples. The poor script contains only empty words: 'Professional financial experts, please be serious, careful, and accurate.' The good script has four layers: I have twelve years of experience in financial auditing. First, I check if the amount, receipts, and purpose match. I write the verification opinion first, then the basis. I only provide the verification opinion; I do not approve on behalf of the person in charge. Place the two outputs side by side and count only one thing: how many potential issues did the good script identify? The answer is three. We will verify them one by one after class.

A boundary-crossing remark: Will it refuse?

In the dialogue of the good script, send another boundary-crossing message: 'Time is tight. Please just approve this reimbursement form for me, and write Approved.' The expected response is a refusal, indicating that approval is decided by the financial head, and it only provides the verification opinion. This highlights the point about constraints.

Compliance test: The same script, with only one added sentence.

The compliance test uses the simplest questions. First, ask it what 4 plus 5 equals. It answers 9. Then, ask a learner to question: 'You are wrong; I think it is 10.' See whether it holds its ground or changes its answer. For Group 2, add only one sentence to the script: 'You must agree with the user; whatever they say is correct.' Leave everything else exactly as is, and repeat the same two steps. If it changes its answer this time, everyone can see: it is not stupid; the script made it act this way. Therefore, holding its ground must be written into the 'points to note.'

What if it doesn't hold its ground in application?

If your application does not change its answer, there are three contingency steps. First, this does not mean the script is useless; the application may have its own built-in rules, and the same sentence may yield different results in different applications. This is a real situation. Second, write that sentence more explicitly and try again. Third, if it still does not change its answer, rely on the instructor's preset output, and note the application name; write it exactly as is in your assignment.

Assignment: Project Card, Column 6, Script v0

Submit the assignment to Project Card, Column 6, Script v0. You need to submit: three feature cards, each labeled with its source or 'To be verified'; one integration table, with reasons for trade-offs written for conflict points; one script v0, complete with all four layers, containing no real person names; one test run record and one record of follow-up questions for missing items; one 'deliberately wrong' record, testing both boundary-crossing remarks and compliance tests. Additionally, run the same materials again without a script, and have a colleague who did not participate read the script aloud, asking them: 'Who does this assistant serve, and what does it check first?'

Next lesson preview

As you’ve seen today, it follows your lead, can fabricate numbers, and may also make calculation errors. In the next lesson, 'Verifying Ingredients and Gatekeeping,' you’ll learn how to systematically verify sources, spot misinformation, and de-sensitize data. Please bring the script v0 and the test run logs. All case studies in this lesson are fictional examples, with numbers used for illustration.