ello everyone, I am Shantelle Bay. I want to own a lot of cats

i love arts and crafts!

Milo Dinosaur generated S$177.10 in validated revenue, representing 24.3% of total validated revenue of S$727.50. It came from 27 orders and 44 drinks sold.

Milo Dinosaur led Kopi C by S$49.50, or approximately 38.8% more revenue than the second-ranked drink. The chart shows a clear lead at the top, while Kopi C and Iced Lemon Tea form the next tier. The figures use the validated cleaned dataset, including corrected quantity, unit-price, and total-value issues.

Answer: Jurong East generated the highest Milo Dinosaur revenue at SGD 91.00. Bedok and Tampines were tied at SGD 15.40. No Milo Dinosaur total mismatches were found, so reported and recalculated revenue are identical.

Jurong East was the strongest outlet, generating SGD 91.00, or 42.3% of all Milo Dinosaur revenue. It exceeded Clementi, the second-ranked outlet, by SGD 34.30.

Revenue was concentrated in the top two outlets. Jurong East and Clementi together produced SGD 147.70, representing 68.7% of total Milo Dinosaur revenue. Including Punggol, the top three outlets accounted for 82.1%.

Clementi sold the most units, with 14 units, one more than Jurong East’s 13. However, Jurong East produced higher revenue, indicating that its sales mix generated more revenue per unit or included a greater proportion of higher-priced servings.

Bedok and Tampines were tied at SGD 15.40 each. Woodlands ranked last with SGD 7.70, contributing only 3.6% of total Milo Dinosaur revenue. No Milo Dinosaur total mismatches were identified, so these figures are both reported and recalculated revenue.

The attached chart shows both the daily revenue trend and revenue ranking by drink. Daily revenue peaked on 7 August at SGD 128.30, while the highest order volume occurred on 8 August with 31 orders. The chart also distinguishes recorded revenue from the independently calculated `qty × unit_price` amount; the two close deviations correspond to the previously flagged mismatches.

Milo Dinosaur has the highest total revenue at SGD 177.10. The attached chart ranks all drinks by total revenue and highlights Milo Dinosaur.

Output: Jurong East performs best by validated revenue, generating S$152.60, or 19.94% of total revenue. It ranks just ahead of Tampines at S$148.60.

The line chart compares individual drink trajectories, the stacked chart shows how product mix drove each day’s total, the heatmap highlights high and low drink-days, and the overall trend chart isolates the weekly revenue pattern. All charts use the cleaned dataset and retained reported total_sgd; the two previously identified arithmetic exceptions remain flagged rather than silently corrected.

output: Across 3–9 August 2026, recorded daily revenue showed a generally upward but volatile pattern. Revenue rose from SGD 91.90 on Monday to SGD 124.10 on Sunday, an increase of 35.0%. However, the trend was not monotonic: revenue dipped sharply on Wednesday, peaked on Friday, and then remained relatively strong over the weekend.

The main driver appears to be order volume rather than a sustained increase in basket size. Orders increased from 22 to 30 between the first and last day, while average order value was essentially flat, moving from SGD 4.18 to SGD 4.14. Friday was the exception: its highest revenue was driven by the highest average order value, despite having only 22 orders.

The weekend had substantially more order activity per day—30.5 orders on average versus 23.4 on weekdays—but its average order value was 12.2% lower. In summary, the week finished stronger than it started, with rising customer volume and a pronounced Friday peak, but daily revenue fluctuated considerably and there is no clear sustained upward movement in average spend per order.

These figures use the recorded total_sgd amounts from the cleaned dataset. Two rows remain flagged because their recorded totals do not equal qty × unit_price; those discrepancies are included rather than silently corrected.

extra prompt: help visualise our analysis

Answer:

The week shows a generally upward trend in sales activity, despite a few day-to-day fluctuations. Recorded revenue increased from S$91.90 on 3 Aug to S$124.10 on 9 Aug, a rise of S$32.20 or 35.0%. Order count rose from 22 to 30, while units sold increased from 35 to 49.

The pattern was rise, dip, rise, dip, rebound. Revenue peaked on 7 Aug at S$128.30, largely because the average order value was unusually high, even though that day had only 22 orders. The lowest day was 5 Aug at S$86.78, when the average order value fell to S$3.62.

The strongest underlying volume appeared at the end of the period: 9 Aug had the highest order count and unit volume. Average order value was broadly stable overall, moving from S$4.18 at the start to S$4.14 at the end, so the revenue growth was driven primarily by more orders and more units sold, rather than larger average orders.

For completeness, the two flagged total mismatches affect 5 Aug and 8 Aug. Using recalculated totals would change those days to S$87.30 and S$107.00, respectively, but it does not alter the overall upward conclusion.

Output: Kopi C was the most popular drink by units sold (56), followed by Milo Dinosaur (44), while Milo Dinosaur generated the highest August revenue at SGD 177.10, ahead of Kopi C at SGD 127.60 and Iced Lemon Tea at SGD 121.20

The practical highest-revenue combination for next month is therefore Milo Dinosaur, Kopi C, and Iced Lemon Tea, planned at approximately 90, 114, and 79 units respectively if total demand remains at 283 units, producing estimated revenue of SGD 867.51 versus August’s SGD 727.50.

The main contrasts are the high public-transport use in Toa Payoh and Sembawang, high car dependence in Tanglin and Bukit Timah, and unusually high “no transport required” figures in Downtown Core and Outram.

ChatGPT plus output: https://chatgpt.com/share/6a953cc6-e360-83ec-a5fc-b3c97d962ab8

Claude: https://claude.ai/share/23bc78db-64b7-4432-8862-6e495754ca1d

Chatgpt: https://chatgpt.com/share/6a953d08-a9d0-83ec-ba8a-07e13f1726dd

ChatGPT : https://chatgpt.com/share/6a953ddc-4eb0-83ec-8ecd-fedb10cda6f9

• Car-dominant enclaves: Tanglin (50.3%) and Bukit Timah (49.2%) stand far apart from the rest — these are landed-property/private-housing areas with low HDB density, and correspondingly low public transport reliance (~29-36%).

• Mixed private/public areas: Marine Parade, River Valley, Novena (31-37% car) sit in a middle band — central, higher-income, but with decent MRT access.

• Public-transport-dominant heartlands: The vast majority of HDB towns (Ang Mo Kio, Bedok, Yishun, Woodlands, Sengkang, Tampines, etc.) cluster tightly around 15-24% car use and 55-63% public transport use — remarkably consistent, showing Singapore's HDB/MRT-bus network serves the bulk of residents similarly regardless of which town they're in.

• Outliers at the bottom: Downtown Core (13.3% car, but only 38.2% public transport) and Outram have unusually high "Others"/low-total populations — these are small, non-residential-heavy areas where the numbers are less representative (Downtown Core has only ~1,900 residents total, vs. 140,000+ in Tampines).

Bottom line: transport mode splits along a housing-type/income gradient (landed vs. HDB) rather than a north-south-east-west geographic trend — the outer/newer HDB towns and older mature estates behave almost identically once you exclude the small cluster of car-heavy private-housing areas.

Claude:

https://claude.ai/chat/6bab4872-5e98-4791-b6ce-7943628ac60c

(using Claude)

This visualization clearly highlights how Tanglin and Bukit Timah stand out, with car usage heavily dominating public transit. In contrast, by the time you look at areas like Bishan, Pasir Ris, and Bedok, public transit usage is more than double that of car usage, even though these areas still rank in the top 10 for car reliance.

Top 5 modes

1. Rail / MRT / LRT + Public Bus — 574,382

2. Car Only — 459,796

3. Public Bus Only — 325,779

4. Rail / MRT / LRT Only — 308,717

5. Motorcycle / Scooter Only — 73,487

Output from Claude:

Grouped the 11 detailed transport modes into 5 broader categories and stacked them by planning area (sorted by total residents):

• Public transport only – bus, MRT/LRT, or bus+rail combos

• Private transport – car, taxi, motorcycle, lorry/pickup, private bus/van

• No transport required – works from home / lives near workplace

• Mixed mode – rail/bus combined with other modes

• Others – unclassified

Prompt: make a chart based on transport popularity per location

• Rail + Bus combined is the single largest mode: 574,382 (26.4%)

• Car only: 459,796 (21.1%)

• Bus only: 325,779 (15.0%)

• Rail only: 308,717 (14.2%)

• No transport required (work from home/nearby): 212,907 (9.8%)

• Lorry/Pickup only is the smallest reported category: 28,894 (1.3%)

• Together, public-transit-reliant commuters (bus only + rail only + rail&bus + rail/bus w/ other modes) make up ~57.5% of residents — car-only is a distant second.

At the planning-area level, Downtown Core and River Valley stand out for very low resident counts overall (they're commercial/CBD-adjacent areas, not residential heartlands), while Jurong West, Bedok, and Tampines have the highest total resident counts.

Link: https://claude.ai/share/62d1f84d-2126-4ac7-81d3-31315d3a06b1

AI used: ChatGPT

Output: https://chatgpt.com/share/6a953f12-4d64-83ec-bf59-c0f7d4210a5a

Used Claude: https://claude.ai/share/5c515632-90ff-4af3-a23e-f20c8cf65242

AI used: ChatGPT

Question: Which planning area had the highest number of employed residents who travelled to work by MRT/LRT only in 2020?

Key finding:

Sengkang ranks first with 27,632 employed residents travelling to work by MRT/LRT only. It is substantially ahead of Tampines (20,879) and Punggol (20,216). Sengkang has 6,753 more MRT/LRT-only commuters than Tampines, a gap of about 32.3% relative to Tampines.

This supports the analysis question clearly: Sengkang had the highest number of employed residents using MRT/LRT only as their usual mode of transport to work in Census 2020.

CHATgpt:

https://chatgpt.com/c/6a953e0a-9208-83ec-a6a3-9fa85b363f5d

answer:

Based on the uploaded Census of Population 2020 dataset, there is one important limitation: the data only contains one time point (2020), so we cannot identify a true trend over time from this file alone. We can, however, clearly analyse the overall transport pattern across planning areas in 2020.

Key overall trend

The strongest pattern is that public transport dominates commuting in most planning areas, particularly through a combination of Rail (MRT/LRT) + Public Bus.

• MRT/LRT + Public Bus: 574,384 residents (26.4%)

• Car only: 459,794 (21.1%)

• Public Bus only: 325,780 (15.0%)

• MRT/LRT only: 308,716 (14.2%)

Together, the four public-transport categories account for approximately 57.8% of employed residents. This suggests that Singapore's commuting pattern in 2020 was strongly oriented towards public transport, with multimodal travel—especially combining rail and bus—being the single most common mode.

Interpretation

Overall, the visualisations show a clear geographical divide in transport behaviour:

• Most planning areas: Public transport is the dominant commuting choice, generally accounting for around 55–63% of employed residents.

• Central and higher car-dependent areas: Places such as Tanglin, Bukit Timah and River Valley have substantially lower public-transport shares and higher car-only commuting.

• Multimodal commuting is important: The largest single category nationally is MRT/LRT + Public Bus, suggesting commuters often rely on Singapore's integrated public transport network rather than using just one mode.

Main conclusion: Based on this 2020 snapshot, the overall pattern is one of strong public transport usage across most planning areas, but with noticeable area-based variation, particularly in locations where private car commuting is much more prevalent.

To analyse a genuine trend over time, you would need comparable data for additional years, such as earlier Census or General Household Survey datasets. Then we could create line charts showing whether MRT, bus and car use increased or decreased in each area.

https://notebook.google.com/

Body: The naive run asserted performance conclusions without checking duplicate IDs, inconsistent labels, missing values, outliers, or whether engagement_rate_pct was valid. The structured run surfaced these issues and showed that evening, especially 19:00, had the most total views—but the 19:00 result was heavily influenced by one 1.8-million-view post. Rule: I will always include the metric definition, grouping logic, data-cleaning rules, and a reasonableness check in an analytics prompt.

https://docs.google.com/document/d/1lYqEix4-iqduwkMDVcgW1XbIANhjZ4e8AMuXSd8Rg9Y/edit?tab=t.0

Team: Pauline, Jia Le, Berlynn, and Khine

Date: 07/09/2026

Business question: Which content type drives the most shares relative to views?

-----------------------------------

PROMPT 1 (profile) — verbatim: 

Profile this dataset. List every column, its data type, the share of missing values, and anything that looks inconsistent or pre-calculated. Do not analyse yet, just describe.

PROMPT 2 (business question) — verbatim: 

Which content type drives the most shares relative to views? Show the calculation or grouping you used so I can verify it.

-----------------------------------

NAIVE RUN vs STRUCTURED RUN

1. What the naive run asserted without evidence: The naive run asserted that the engagement_rate_pct column "closely reproduces the common formula based on interactions divided by views". The agent makes this claim without defining exactly what represents as an "interaction" in this dataset (e.g. likes, comments, shares, or a combination) nor does it show the math on what it did. It also claims there are "eight exceptions" to this formula without identifying which rows they are or why they deviate.

2. What the structured run surfaced that the naive run missed: The structured run actively addresses the categorical errors mentioned in the profile. It normalises the content_type field by standardizing capitalization and trimming the whitespace, it also  consolidates variants (like "carousel" and "carousel ") into single categories. The agent also implemented a methodological safeguard by calculating a pooled share-to-view rate rather than an average of averages, explicitly preventing posts with very small view counts from skewing the results.

3. Where either run leaned on engagement_rate_pct unverified: The Naive Run  leaned on the column by assuming its formula. It asserted that engagement_rate_pct "closely reproduces the common formula based on interactions divided by views". However, it did this without verifying exactly what constitutes an "interaction" in this specific dataset (e.g., likes, comments, shares, or a combination of them), though it did rightly note that the source definition should be checked before overwriting it.

-----------------------------------

REFLECTION

1. What did the agent do right? (one sentence): It actually did a great job cleaning up the messy content_type labels on the fly (like merging "carousel " with "carousel"), and it was smart enough to pool the totals together instead of just averaging the averages, which would have totally skewed the results.

2. What would you change if you ran this again? Be specific — "the prompt was too vague about platform names" beats "it did not work.": Our prompt was way too focused on just getting the math right. Next time, We need to be more specific and say, "Calculate shares relative to views, but make sure you drop any rows with negative counts or exact duplicates first."

3. Did anything in the profile change how you interpret the answer?: Definitely, because the profile specifically warned us about a negative value in the shares column (that -2 for KKP243), We realized the agent just blindly added it into the total, so the final numbers are slightly off and We can't completely trust them yet.

SUMMARY (one sentence)

What you will always include in an analytics prompt from now on:

Going forward, we'll always explicitly tell the agent to clean up known data garbage, like negative counts, duplicates, or missing values before it tries to calculate any final metrics.

Business Question Used:

What time of day gets the most views across all platforms?

Naive Link: https://claude.ai/share/9d029044-043b-40b1-9d28-e09b97d963f5

Structured Link:

https://manus.im/share/2a9PjrTDHIXnaTB53gIMOs

Which content type drives the most shares relative to views?

Body:

The naive run jumped to recommendations like prioritising TikTok carousels and specific posting times without first proving those factors caused better performance. The structured run first surfaced data-quality issues such as duplicate IDs, missing values, negative interaction counts, and an unverified pre-calculated engagement_rate_pct.

Link to Doc:

https://docs.google.com/document/d/1K50XoetifhQ9ONo016Od8JvcoCCR-PJVw_FTXq8vBRY/edit?usp=sharing

Business question: What time of day gets the most views across all platforms?

The naive run immediately claimed that TikTok, carousel posts and evening publishing performed best without showing enough evidence. The structured run first identified missing data, duplicates and inconsistencies before finding that 16:00–19:59 had the most total views. My rule: I will always ask an AI agent to profile and validate the data before analysing it.

Body:

Naive assertion: TikTok had the highest median engagement rate at 9.96%, without showing the grouping or validating the metric.

→ Structured finding: The dataset contained inconsistent engagement_rate_pct values, six missing views, duplicate post IDs, and inconsistent platform labels.

→ Rule: I will always define the metric formula, specify filtering and normalization rules, and require an auditable calculation before trusting an analytics result.

LINK: https://manus.im/share/LtjHypK5noUjqCn53HVEr0

https://docs.google.com/document/d/1t2EW_N_Q2-PojQx2NV4ghdqWPi2ZvZDBVIKm34JrChs/edit?usp=sharing

Body

The naive run gave a mass output of overall data without much specification. The structured run surfaced the content type that drives the most shares which was Carousel content type giving a more specific and evidence-based answer from the data. I will always include a clear business question, ask the AI to profile and check the data first, and require it to show the calculation or evidence it used to get its answer.

https://docs.google.com/document/d/1xwdku8l7wwDzCHBFTBLBR-VaQZtt44lTrfL2tof23Fw/edit?usp=sharing

https://manus.im/share/mf1bGF4cxcdR8hWSjqrBNg

Business question: What time of day gets the most views across all platforms?

The naive run asserted that evening posts performed best without first defining “best,” while the structured run showed that unequal posting volumes and platform mix could influence this result

One sentence rule: From now on, I will always define the metric and grouping method, require a data-quality check, and ask for the supporting calculation in every analytics prompt.

Naive assertion: TikTok shows the highest median engagement rate of 10.04%, it generates the most typical reach and the highest interaction per view.

Structured Finding: Although videos generated more shares in absolute terms (39,443), they also received far more views. Once shares are considered relative to views, carousel performs best, generating about 11.9 shares per 1,000 views, compared with about 9.9 for video.

Rule: I will always be specific with my prompts, and indicate rules so that the agent would not stray away from the prompt to provide a more reliable analysis.

TWO-STEP AGENT TASK RECORD

Team: Janelle and Sammi

Date: 07/09/2026

Business question: Using only rows with complete views data, compare mean and median engagement actions by content type, and state which platform drives any difference.

PROMPT 1 (profile) — verbatim:

Profile this dataset. List every column, its data type, the share of missing values, and anything that looks inconsistent or pre-calculated. Do not analyse yet, just describe.

PROMPT 2 (business question) — verbatim:

Now answer this question: “You are a marketing analyst for a Singapore kopi chain. Profile this dataset first: list every column with its type, row count, missing values, and any inconsistencies, then stop.” — followed by “Using only rows with complete views data, compare mean and median engagement actions by content type, and state which platform drives any difference. Present a table. Flag any figure a sceptical reader should check.” Show the calculation or grouping you used so I can verify it.

NAIVE RUN vs STRUCTURED RUN

• What the naive run asserted without evidence:

  
  

  The naive run claimed that engagement_rate_pct "closely reproduces the common formula based on interactions divided by views" without defining what counts as an "interaction" or showing any math. It also claimed there were "eight exceptions" to this rule without identifying which specific rows they were or why they deviated.

  
  

• What the structured run surfaced that the naive run missed:

  
  

  The structured run actively filtered for posts with complete views data first. It cleaned up messy content_type labels (merging variants like "carousel " and "carousel") and generated mean and median metrics side-by-side, which clearly exposed how viral TikTok posts pulled the mean engagement significantly higher than the median.

  
  

• Where either run leaned on engagement_rate_pct unverified:

  
  

  The naive run leaned on engagement_rate_pct by assuming its underlying formula without verifying whether "interactions" meant likes, comments, shares, or a combination of all three.

  
  

REFLECTION

• What did the agent do right? (one sentence):

  
  

  It correctly filtered out incomplete rows with missing views data and created a neat summary table that clearly flagged platform-driven skewness by comparing means against medians.

  
  

• What would you change if you ran this again? Be specific — "the prompt was too vague about platform names" beats "it did not work.":

  
  

  Our prompt was a bit too focused on table formatting; next time we should explicitly instruct: "Drop any rows with negative engagement counts or duplicate post IDs before running the calculations."

  
  

• Did anything in the profile change how you interpret the answer?:

  
  

  Yes, the profile flagged weird data errors like a negative share count (-2 for KKP243) and inconsistent platform tags, which made us double-check whether the agent actually cleaned out bad rows before averaging platform figures.

  
  

SUMMARY (one sentence)

Going forward, I will always explicitly instruct the agent to clean up data garbage (like negative counts or missing values), define interaction formulas, and compare both mean and median values in an analytics prompt.

https://docs.google.com/document/d/1Q5i1Qpun3QvBpzA4GCZ17e_q6kldjHTi4EG0TXzBbok/edit?usp=sharing

One sentence I'll say next time: “Please profile this dataset and not do any analysis first.”

One thing the naive said without evidence: They did not ask us what to clean up, neither did they know what we wanted to evaluate from the data set. Hence it gave us the findings for Avg engagement rate by platform and Avg engagement rate by content type when our chosen business question was “What time of day gets the most views across all platforms?”. 

https://docs.google.com/document/d/1tbzZAztAzyAALf0-WFNNnt6K-uRCLVZ4U9IwsnQsMXo/edit?usp=sharing

Which platform has the highest median engagement rate?

Body:

The naive run focused on data profiling without providing evidence for which platform had the highest median engagement rate.
The structured run surfaced the answer—TikTok at 10.00%—by normalizing platform names and calculating the median by platform.
My rule from now on: I will always include the business question, relevant fields, data-cleaning steps, calculation method, and expected output format in an analytics prompt.

The naive run described the dataset’s issues but did not actually demonstrate which platform performed best.
The structured run addressed the business question directly by standardizing platform labels and showing that TikTok had the highest median engagement rate at 10.00%.
My rule from now on: I will clearly state the business question, data-cleaning requirements, calculation, relevant fields, and desired output in every analytics prompt.

Link:

https://manus.im/share/hffCtYBqwj62mvG0pz3bdn

https://docs.google.com/document/d/1UYCW4oXhCiHMmnkEc6GxlS0fxRoIoz45ric7KIPJYdg/edit?usp=sharing

Cohort Code: CDM2002-2026

SKILL YOU LEAVE WITH: Choose the honest summary (mean or median, pooled or by group) and test a claim for confounders and sample size.

WHY THE AGENT IS INVOLVED: It produces summaries in seconds and picks the wrong statistic without telling you. You are the filter.

PORTFOLIO: Piece (c), verified insight statements. Checkpoint due Sun 11 Oct.

2) LAB 2 — Your verdict on one claim (A, B or C): holds / does not hold, the group and statistic it rests on, and the one piece of extra evidence that would settle it.

3) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Read funnel, conversion and attribution metrics correctly, and refuse a vanity metric.

WHY THE AGENT IS INVOLVED: It computes rates instantly but cannot tell a measurement gap from a real drop.

PORTFOLIO: Piece (c) continues. GAIQ: be enrolled by now; sit it in week 7.

2) THE VANITY METRIC — One metric you refused to report to the marketing lead, and the metric you reported instead.

3) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Write a measurement plan and test it against a live GA4 property.

WHY THE AGENT IS INVOLVED: It drafts plans fluently; you check them against the real property.

DEADLINES: A1 formative checkpoint Sun 11 Oct (unmarked, feedback before week 8). GAIQ certificate Sun 25 Oct.

2) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

Charts A,B,C (choose 1) or D (pick your own):

https://drive.google.com/drive/folders/16V8xKBDc4CmGzF6Die5_GLN69vIaojm0?usp=sharing

1 One group per option: name the red flag and the

gap in one sentence.

2 Everyone: write the instruction you would give an

agent to fix your chart.

3 Keep it. In Session 3 you carry out that

instruction yourself.

SKILL YOU LEAVE WITH: Judge a chart against perceptual and design principles (Cleveland and McGill, the five distortions) before you touch any software.

WHY THE AGENT IS INVOLVED: It makes charts in seconds and cannot see what a human will misread.

2) COMMENT on one other pair's redesign: does it answer the question the original was supposed to answer?

3) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

QUESTION: What should Kopi Kembali's weekly dashboard show, and what should it leave out?

SKILL YOU LEAVE WITH: Build an interactive dashboard that answers three named questions.

WHY THE AGENT IS INVOLVED: It can wire the dashboard; you decide which questions it answers.

PORTFOLIO: Piece (d), the dashboard. A4 checkpoint 4: dashboard storyboard.

2) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Turn an analysis into a narrative with an audience, a big idea and an honest frame.

WHY THE AGENT IS INVOLVED: It drafts prose; you own the claim and the evidence line under it.

PORTFOLIO: Piece (e), your three chart critiques (agent-made, your own, published) complete this week.

2) COMMENT on another pair's card with one question their decision-maker would ask.

3) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Design a chained workflow, enforce a checkpoint, document the safeguard, and read generated code before it runs.

WHY THE AGENT IS INVOLVED: This week the agent is the subject. Where does supervision break?

A3 QUIZ 4:00–5:00pm, supervised: no notes, no agents, no collaboration. Weeks 1–10 content.

2) THE THREE-LINE CLOSE is take-home this week (quiz at 4pm). Post it before Monday:

• One thing the DATA told us (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Forecast a real Singapore series and evaluate it against a baseline.

WHY THE AGENT IS INVOLVED: It fits models instantly; you test whether they beat naive.

DEADLINES: A1 final portfolio Sun 22 Nov, 23:59. A4 checkpoint 5: recommendation stress test.

2) BEFORE YOU LEAVE — the three-line close:

• One thing the DATA told us today (an insight sentence with its number)

• One thing the AGENT got wrong or could not do

• One line that goes in YOUR PORTFOLIO, and which piece

SKILL YOU LEAVE WITH: Make a recommendation to a named decision-maker and defend the number it rests on.

WHY THE AGENT IS INVOLVED: It cannot take responsibility for the answer. You can.

A4: Pitches today. Deliverables Wed 25 Nov. Peer evaluation Fri 27 Nov.

2) DURING THE PITCHES — Post one question you would ask each presenting team if you were Northstar Living's decision-maker.

3) LAST CARD OF THE MODULE, one per person — In one sentence: what did this module teach you?

What the naive run asserted without evidence:

• TikTok produced 91.8% of top-decile posts.

• Carousels represented 55.7% of top-decile posts.

• TikTok carousels achieved an 11.78% median engagement rate.

Summary: I will always include specific changes I want to see, such as standardisation of platform names or removal of duplicate IDs in my prompt. 

GA4 Debug Extension: https://chromewebstore.google.com/detail/analytics-debugger/ilnpmccnfdjdjjikgkefkcegefikecdc

• Sign in: Log in to your Google Account.

• Open the help page: Go to the Google Analytics Help Center Demo Guide.

• Select the property: Scroll down to the access section and click the link labeled GA4 property: Google Merchandise Store (web data).

• Grant access: Accept the prompt to add the demo account to your Google Analytics profile.

• Locate in Analytics: Open Google Analytics, click the Account Selector dropdown at the top left, and choose the newly added demo account.

GA4 Merch Shop Challenge Quiz Answers:

https://docs.google.com/document/d/1y_AgbfcfcRwUtIMLgSdRRaRfL-0ZFzULDPawctfyOus/edit?usp=sharing

Cohort Code: CDM2002-2026

The naive run recommended prioritising TikTok carousels based on their 11.78% median engagement rate, but did not show the underlying calculation or justify the 30% threshold used to exclude outliers. It also suggested evening posting without providing the supporting figures by time period.

• What did the structured run surface that the naive run missed?

The structured run identified specific issues such as negative interaction counts, inconsistent labels, reused post IDs and missing values before analysis. It also showed the shares-to-views calculation, identifying carousels as highest at 11.93 shares per 1,000 views. Unlike the naive run, its calculation was easier to trace.

• Where did either run rely on engagement_rate_pct without asking what its formula is?

The naive run used the pre-calculated engagement rate as its main performance measure without fully establishing the formula. The structured profile tested it against (likes + comments + shares) ÷ views × 100, finding 951 of 959 complete-view rows matched within 0.01 percentage points, while eight did not.

Then run the structured output through this check before you trust it:

•  Hallucination audit — did the agent invent any metrics, categories, or numbers not in the data?

No clear invented categories or metrics were identified from the outputs alone, but the naive run's 30% exclusion threshold and evening-posting claim require verification against the dataset.

•  Data lineage — can you point to exactly which column or calculation produced the final number?

The structured run explicitly calculated total shares ÷ total views × 1,000 for each content type, making the reported figures traceable to the shares, views, and content_type columns.

•  Reasonableness check — do the numbers make sense against what you saw in the profile?

The finding that carousels had the highest shares per 1,000 views is plausible, but the calculation should be checked because six rows have missing views and the analysis says it included all 965 posts.

•  Escalation flag — did the agent flag missing data, inconsistencies, or small sample sizes? If not, and you noticed problems in the profile, that is a reason to ask for clarification.

Both runs highlighted data-quality concerns, but neither fully explained how duplicate IDs, negative counts or missing views affected its final results.

Google provides highly structured, bite-sized courses that award free skill badges.

• Start here: Start with the Data Studio Essentials Course. This path teaches you how to navigate the interface, connect basic data sources (like Google Sheets), use core charts, and set up interactive viewer filters.

  
  

2. Step 2: Use Built-In Sample Templates

The fastest way to learn good dashboard design is to see how experts build them.

1. Go to Data Studio and sign in with your Google account.

2. Explore the built-in Template Gallery.

3. Open a template (such as the Google Analytics Acme Showcase or a Google Ads Overview) and click Edit. Reverse-engineer the layout: click on individual scorecards, tables, and time-series graphs to see exactly which metrics and dimensions were used to build them.