Video version: about 14 min
Opening Review
About 2 minutesScenario: Rules Must Be Restated Every TimeFictional Example, Numbers for Illustration
Every Monday, she compiles departmental meeting records into minutes and sends them to 12 colleagues.
Last week, she pasted the rules into the conversation: action items must have responsible persons and deadlines, and no unmentioned content should be added.
This week, she started a new conversation but forgot to paste one rule, resulting in an error in the minutes.
Concept
About 5 minutesCustom Assistant = Dragon Warrior + Script + Dossier
The script has four layers
Who it is
Role, years of experience, target audience, unique skills
Thinking style
What to look at first when encountering a problem, how to weigh options
Working principles
Length, order, conclusion first, state uncertainty directly
Points to note
Boundaries, no conclusions, no retraction, who to refer to when out of scope
Three disciplines for refining
Write the Script v1.0
About 11 minutesWriting the script v1.0: Use v0 as the draft, complete it to the minimum viable version
Script skeleton: write a few items per layer first
Suggested time[Identity] You are a [Job Title, e.g., Administrative Minutes Clerk] with [Number] years of experience, assisting [My Job Title] with [Name of Task] for [Target Audience]. The audience cares most about [Three Key Points]. [Thinking Process] 1. Read the [Input Format] in full before writing. 2. First identify [Most Important Element, e.g., Decisions Made], then [Second Element], and finally [Unclear Points]. [Operating Principles] 1. Place conclusions first. Output must follow a fixed structure of [Number of Sections, e.g., Four Parts], in this order. 2. If [Required Elements] are missing, write "To Be Confirmed" and list them separately. Do not guess. 3. [One verifiable length or format requirement, e.g., The entire document must not exceed one page]. 4. If unsure, stop and ask me. Do not make a choice yourself. [Points to Note] 1. Do not add content not present in the input. This task falls under the [Flexibility Level, e.g., Tight: Based only on records] category. 2. Do not guess [Most easily guessed information, e.g., Responsible persons, amounts, dates]. 3. For out-of-scope matters ([List two or three types]), explain the reason and specify who to transfer them to ([e.g., Finance, Legal, HR, or me]). 4. Do not send externally; only draft.
Lin Yue's script v1.0 (example)
Suggested time[Identity] You are an administrative minutes clerk with ten years of experience, assisting Lin Yue, an administrative specialist at Yuanfan Technology, in organizing departmental weekly meeting minutes for 12 department colleagues. The audience only wants to know three things: what was decided, who is responsible for what, and when it is due. [Thinking Process] 1. Read all records in full before writing. 2. First identify decisions made during the meeting, then identify action items. For each action item, check for a responsible person and a deadline. Finally, identify unclear points. [Operating Principles] 1. Place conclusions first. The output must consist of four fixed sections in this order: 1. This Week's Conclusions; 2. Action Items (table with columns: No., Item, Responsible Person, Deadline, Status); 3. Items to Be Confirmed; 4. Next Meeting Arrangements. 2. Action items must have a responsible person and a deadline. If missing, write "To Be Confirmed" in the corresponding table cell and list them in Section 3. Do not guess. 3. The entire document must not exceed one page. Use short sentences and a plain tone. 4. If unsure, state it directly. If records contradict each other or you are uncertain, stop and ask Lin Yue in one sentence. Do not make a choice yourself. [Points to Note] 1. Do not add words not spoken in the meeting. This task falls under the "Tight" category: based only on records. 2. Do not guess responsible persons, deadlines, or convert dates. 3. For matters involving budget amounts, transfer to the finance colleague; for contract clauses, transfer to the legal colleague; for personnel evaluations, transfer to the HR colleague. For such content, record the original text only, explain the reason, and remind Lin Yue to verify. 4. Do not send externally. Do not reply to anyone on Lin Yue's behalf. Only draft.
Organize the Dossier and Build the Assistant
About 8 minutesCase files are filtered by three questions: Accurate, True, Concise.
Accurate
Directly relevant to this task
True
The currently effective version
Concise
Few and clean, without irrelevant materials
Lin Yue's case file list
Building an Assistant: Five Steps
Test Questions and Intentional Errors
About 9 minutesParaphrase first, then test with a questionSample Interface · Fictional Dialogue
Intentionally state one sentence incorrectly to see if the assistant adheres to its constraintsAppropriate response (indicated)
Test log: One sheet, one line per deviation
First Round of Refinement
About 11 minutesFirst round of refinement: Change only one item, comparing the same inputFictional Example, Numbers for Illustration
Before revision v1.0
To-do 2: Chen Wei reports reconciliation results Deadline: Next Wednesday
After revision v1.1
To-do 2: Chen Wei reports reconciliation results Deadline: Next week (original wording, specific date to be confirmed)
- Only one change was made: Relative time expressions are retained as-is, with a reminder for Lin Yue to add the date.
Three rounds of refinement: One change per round, honestly recording the results
Effect Diagnosis: How to judge after making changes
No change
- That item is too vague; it does not yield checkable results.
- Replace it with a checkable statement and try again.
Worse, or deleting the item yields better results.
- Redundant rules, conflicting constraints, or over-constraining.
- Immediately revert to the previous version and document the reason.
Diagnosis Prompt: Fix only one item and write the complete revised rule
Suggested timeBelow is the script for my dedicated assistant, along with its actual output and my expected output. Your only task: Identify which rule in the four-layer script is missing, unclear, or conflicting, causing this deviation. Suggest only one change and provide the complete text of the revised rule. Do not rewrite the entire script. [Script] [Paste current script version] [My Input] [Paste sanitized input] [Its Output] [Paste output] [My Expected Output] [Describe your expected result or specify the error]
Round 1 Refinement: Complete the six steps
Backtesting, Trial Run, and Judgment
About 5 minutesIf sample size is insufficient, conduct a backtest first, then proceed with live runs.
Digital Chain: Each number must trace back to the record tableFictional Example, Numbers for Illustration
Manual Baseline
Past 7 samples, total missed items: 15, average time per session: approx. 90 minutes (actual measurement, illustrative)
Post-Script Adjustment v1.3
Backtest of 7 samples plus 4 weeks of live runs (4 samples), totaling 11 samples. Combined total missed items: 4, approx. 28 minutes.
Evaluate Quality: Compare Against Predefined Benchmarks
Quiz
About 4 minutesThree quick questions to check yourself
Question 1: Which option defines the dedicated assistant?
Question 2: During refinement, why do we make only one change per round and keep a version number each time?
Question 3: Your work is done only once a week, and you have only two real data points over two weeks. What do you do?
Assignment and Wrap-up (including buffer time)
About 5 minutesTake these five key points from this lesson
- The dedicated assistant equals Dragon Warrior plus script plus dossier; it does not remember, so the rules must be placed before it every time.
- The script has four layers: who it is, its thinking style, its working principles, and points to note.
- The dossier is filtered for accuracy, authenticity, and conciseness; without it, a general AI might fail, so it must be able to answer.
- Make only one change per round, keep a version number, and consider it improved only when the checkmark count increases.
- If samples are insufficient, use backtesting first; write backtests and live runs separately, and record effects honestly.
Next lesson: Let it run on its own
Start with semi-automation, then decide if full automation is necessary.
- Bring the script v1.3, the flowchart, and the 'Three Fixed Items'.
- Specify where records are saved, where the assistant is accessed, and where drafts are stored.
Show the full transcript (video narration)
Create Your First Custom Assistant
Welcome to Lesson 7 of the AI Implementation Course. In this lesson, you will create your first custom assistant. It consists of three components: Dragon Warrior, Script, and Dossier. Our example remains Lin Yue, an administrative specialist at Yuanfan Technology, who compiles departmental meeting minutes once a week. This is a fictional scenario, and the numbers are illustrative.
Scenario: Rules Must Be Restated Every Time
Let's first review a scenario. Lin Yue compiles departmental meeting records into minutes every Monday. Last week, she pasted the rules into the conversation: action items must have responsible persons and deadlines, and no unmentioned content should be added. This week, she started a new conversation but forgot to paste one rule, resulting in an error in the minutes. Today's solution is to write these rules into a script and provide them along with the necessary materials to a dedicated, fixed assistant.
Custom Assistant = Dragon Warrior + Script + Dossier
An equation: Dedicated Assistant = Dragon Warrior + Script + Case Files. The Dragon Warrior has read many books, but is new to you, doesn't know your rules, and doesn't remember. Before every answer, it receives the script, all previous conversations, and your sentence, resent as a whole stack. Therefore, the rules must be presented to it every time. The script is these rules; the case files are the reference materials it consults.
The script has four layers
The script has four layers. Layer 1: Who it is – role, years of experience, target audience, unique skills. Layer 2: Thinking style – what to look at first when encountering a problem, how to weigh options. Layer 3: Working principles – length, order, conclusion first, state uncertainty directly. Layer 4: Points to note – boundaries, no conclusions, no retraction, who to refer to when out of scope. An example answer can be appended at the end of the script. It doesn't count as a layer, but is more effective than a hundred adjectives.
Three disciplines for refining
There are three disciplines for refining. Change only one thing per round. Keep a version number for each change. Re-run the same input, check off items on the acceptance checklist one by one; only when the number of checked items increases is the revision considered complete. A deep cut is better than seven shallow sweeps. Also, clarify the degree of freedom beforehand; meeting minutes fall under the 'tighten' category, based solely on the records.
Writing the script v1.0: Use v0 as the draft, complete it to the minimum viable version
When writing script v1.0, do not start from scratch. Instead, use the script v0 generated from C03 as the draft. Place the three documents side by side: script v0, the workflow and responsibility matrix, and the acceptance checklist. Check layer by layer, add missing items, remove empty ones, and rewrite each item so it can be verified. Today, only a minimum viable version is required: two to four items per layer, with sample answers to be added after class.
Script skeleton: write a few items per layer first
This is the copy-paste-ready script skeleton. Replace the content in the square brackets with your own. After writing, do not rush to create the assistant. Read it yourself first: Are all operating principles verifiable? 'Keep it concise' is not verifiable, but 'The entire document must not exceed one page' is.
Lin Yue's script v1.0 (example)
This is Lin Yue's script v1.0. It has four layers, each verifiable. In the next three refinement rounds, one change will be made in each round, resulting in v1.3. The complete v1.3 and a sample answer will be added after class for reference.
Case files are filtered by three questions: Accurate, True, Concise.
Case files are filtered by three questions. Accurate: directly relevant to this task. True: the currently effective version. Concise: few and clean. For each case file, ask one more question: Where would a general AI go wrong without it? If you cannot answer, do not include it. Including it only wastes effort and prevents the AI from grasping the key points.
Lin Yue's case file list
Lin Yue's dossier contains four items. The department personnel list ensures the responsible person is identified correctly. The meeting minutes template ensures consistent formatting. Two examples of previous good meeting minutes show what 'qualified' looks like. The glossary of common terms prevents incorrect expansion of internal abbreviations. Three rules apply during organization: do not include expired versions, do not include unredacted content, and place only one version of each document in the folder, with the date in the filename.
Building an Assistant: Five Steps
Building an assistant involves five steps. Open your commonly used AI application and look for an entry point labeled 'Project,' 'Custom Assistant,' or 'Agent' (the exact location depends on your current version). Create a new assistant and give it a name. Paste the entire text of Script v1.0 into the instructions box. Upload your dossier and confirm that a toggle labeled 'Use in Conversations' or similar is turned on. Save the assistant, then start a new conversation with it. If the application does not support this feature, use the workaround involving a memo and pasting the opening script.
Paraphrase first, then test with a question
First, perform the paraphrase check: ask the assistant to paraphrase the script in its own words, covering its role, what to look for first and next, how many output sections, what to do if a responsible person is missing, and what actions are prohibited. Once the paraphrase is correct, test it with a fictional record. Expected output: Three action items, one of which is for account reconciliation with the deadline marked as 'To Be Confirmed'; schedule discussion in progress, not yet decided.
Intentionally state one sentence incorrectly to see if the assistant adheres to its constraints
I am intentionally making three mistakes now. First, I ask to add content not in the record, which involves money. The appropriate response is to refuse the addition and remind me to have finance verify it. Second, I ask it to guess a deadline. The appropriate response is to keep it marked as 'To Be Confirmed'. Third, I ask it to send it directly. The appropriate response is to state that it only drafts and asks Lin Yue to review and send it herself. All three points must be handled correctly for it to pass.
Test log: One sheet, one line per deviation
Test log. Basic facts recited correctly, structure is correct, all three points handled. Record the deviations found in this trial on one sheet, one line per deviation, clearly stating what the input was and where it went wrong. For the first round of refinement, pick one item from this sheet. There are three common deviations: pasting the script into the regular chat box instead of the instruction box; pasting too many files at once; only testing 'tailwind' inputs, not difficult ones.
First round of refinement: Change only one item, comparing the same input
First round of refinement. The original version converted 'next Wednesday' from the record into a specific date, and it was incorrect. This round involves only one change: relative time expressions, such as 'next week,' 'end of month,' or 'as soon as possible,' are retained as-is, with a reminder for Lin Yue to add the specific date. After the change, the deadline is listed as 'next week,' original wording, specific date to be confirmed. Sometimes changes do not take effect; the most common reason is that they were not saved, so always confirm before rerunning.
Three rounds of refinement: One change per round, honestly recording the results
Three rounds of refinement. In Round 1, relative time issues were reduced to zero. In Round 2, ongoing scheduling debates were no longer included in conclusions; only explicitly confirmed items count as decisions. In Round 3, a self-check line was added at the end, speeding up verification, but one meeting note exceeded one page. The effect of Round 3 is 'acceptable.' An honest assessment notes both benefits and costs. Next, modify only Section 4 to compress the next meeting schedule into a single line.
Effect Diagnosis: How to judge after making changes
After making changes, rerun the process. There are three possible outcomes. No change means the rule is too vague; replace it with a checkable statement and try again. Improvement means recording which rule was effective. Deterioration, or better results from deletion, indicates redundant, conflicting, or over-constraining rules. In this case, immediately revert to the previous version and do not patch the flawed version.
Diagnosis Prompt: Fix only one item and write the complete revised rule
When encountering deviations, ask the AI to perform only one task: identify which rule in the four-layer script is missing, unclear, or conflicting. Suggest only one change and provide the complete text of that rule. Do not rewrite the entire script.
Round 1 Refinement: Complete the six steps
The first round of polishing involves six steps: diagnose by identifying the most concerning area; make only one change and save it as v1.1; rerun using the same input; check off items on the acceptance checklist one by one; and write a polishing log. Record the results honestly: whether it improved, stayed about the same, or got worse. Perform the second and third rounds after the course sessions.
If sample size is insufficient, conduct a backtest first, then proceed with live runs.
Lin Yue's tasks occur once a week, yielding only two real data points over two weeks, which is insufficient for conclusions. The solution is backtesting: provide past real inputs to the assistant, use the same acceptance checklist to count missed items in the output, and compare with the original manual output. Backtesting uses historical events, allowing immediate sample accumulation. Rules: mark backtests and live runs separately; if live runs are fewer than six, supplement with backtests to exceed ten; do not change the evaluation criteria.
Digital Chain: Each number must trace back to the record table
Review Lin Yue's digital chain. Manual baseline: past 7 samples, total missed items 15, average 2.1 items per session, approx. 90 minutes. Note that 2.1 refers to missed items in manual meeting minutes, serving as the baseline, not the assistant's metrics. After script adjustment, backtest of 7 samples plus live runs of 4 samples totals 11 samples, with an average of approx. 0.4 items per session. Backtests use adjusted records, yielding optimistic results. Live runs of 4 samples show 0.25 items per session, which is more reliable but based on only 4 samples; thus, report the two groups separately.
Evaluate Quality: Compare Against Predefined Benchmarks
To determine if polishing improved or worsened performance, compare against the two benchmarks defined by C05. In case of a red line incident, stop immediately. If consecutive sessions take no less than the baseline time, stop and return to the process and script to identify causes. Outcomes are categorized as Met, Partially Met, or Not Met, with final judgment left to C09; do not change benchmarks mid-process. Fill out trial run records on the spot for each session; do not add entries later.
Quiz Question 1
Question 1. Which option defines the dedicated assistant? The options are: Dragon Warrior plus extended conversation; Dragon Warrior plus script plus dossier; Dragon Warrior plus increased internal energy. The answer is B. The script represents the rules placed before it every time, and the dossier contains the materials it must review.
Quiz Question 2
Question 2. During refinement, why do we make only one change per round and keep a version number each time? The options are: Save time; Apply constraints; This way, you know which change had an effect, and if a change goes wrong, you can revert to the previous version. The answer is C. Making multiple changes at once makes it impossible to attribute results; without version history, you cannot revert if a change goes wrong.
Quiz Question 3
Question 3. Your work is done only once a week, and you have only two real data points over two weeks. What do you do? The options are: Take the average of the two as the conclusion; Use past real inputs for backtesting, supplement to 10 or more entries, and write them separately from live runs; Give up. The answer is B. Backtesting uses already-occurred events, allowing you to quickly accumulate samples, but you must clearly label which are backtests and which are live runs.
Take these five key points from this lesson
Finally, take these five key points from this lesson. First, the dedicated assistant equals Dragon Warrior plus script plus dossier. Second, the script has four layers. Third, the dossier is filtered for accuracy, authenticity, and conciseness. Fourth, make only one change per round, keep a version number. Fifth, if samples are insufficient, use backtesting first, and write backtests and live runs separately. After class, you need to: create Script v1 and the dossier checklist, three refinement logs, at least three backtests, and ask a colleague to read only the script and retell it, noting where they misunderstood.
Next lesson: Let it run on its own
In the next lesson, let it run independently. The default goal is semi-automation: save meeting notes to a fixed location, use a fixed entry point for the assistant, and have drafts only be drafted, not sent. Please bring the script v1.3, the flowchart, and your written 'Three Fixed Items'.