Question Classification
Maps author's natural-language what-if question to the right W-mechanism (W1-W6) or the honest-limitation path (parameter-only).
Classification rules
Apply in order. First match wins.
Rule 1. Single-value reference
Match signals (locked exact-match list, case-insensitive). "line," "target," "mark," "break-even," "breakeven," "threshold," "benchmark," "goal."
Example. "Add a line at $1M so I can see which months hit target."
Classification. W1 reference line.
Rule 2. Range reference
Match signals (locked exact-match list, case-insensitive). "band," "range," "zone," "corridor," "between," "tolerance."
Example. "Shade the forecast corridor on the trend."
Classification. W2 reference band.
Rule 3. Projection or extension
Match signals (locked exact-match list, case-insensitive). "project," "extend," "extended," "where will this be," "if we keep going," "at this rate," "at the current rate."
Example. "Project this sales trend out six months."
Classification. W3 trend line extension. Always flag as projection, not prediction (HC-6).
Rule 4. Membership-shift
Match signals (locked exact-match list, case-insensitive). "who would be in," "which would move to," "members of," "top-n if," "cutoff," "move into," "fall below," "fall out of."
Example. "Which accounts would move into the top-20 if they grew 15%."
Classification. W4 set-based scenario.
Rule 5. Side-by-side overlay
Match signals (locked exact-match list, case-insensitive). "overlay," "compare to," "actual vs," "side by side," "on the same chart as."
Example. "Show actual revenue vs. what revenue would be at target."
Classification. W5 dual-axis overlay.
Rule 6. Baked-in hypothetical as a reusable field
Match signals (locked exact-match list, case-insensitive). "a field for," "a calc for," "a measure that," "add it to the data," "create a calc," "create a field." Also matches the word "if" co-occurring with a formula shape ("if every customer spent 20% more").
Example. "Revenue if every customer spent 20% more."
Classification. W6 scenario calc. Handed off to calc-writer.
Rule 7. Parameter-authoring-required (honest limitation)
Match signals (locked exact-match list, case-insensitive). "slider," "dropdown," "let the user pick," "when i change the input," "interactive scenario," "user-selected value," "let the user choose."
Example. "Let the user pick a growth rate from 0 to 50% and update the chart."
Classification. Parameter-only. Emit honest-limitation verdict (HC-3) with manual Desktop parameter path. State plainly that parameter authoring is weak in the current agent and recommend manual Desktop creation.
Rule 8. Monte Carlo or probabilistic
Match signals (locked exact-match list, case-insensitive). "simulate," "probability," "distribution of outcomes," "monte carlo," "random."
Example. "Simulate revenue outcomes if churn is between 5 and 15%."
Classification. Out of scope. Emit honest-limitation verdict. Recommend external modeling (R, Python, Excel) and import results back into Tableau.
Rule 9. Causality or prediction (not supported)
Match signals (locked exact-match list, case-insensitive). "cause," "predict," "forecast with," "confidence," "machine learning prediction."
Example. "Predict next quarter's revenue with 95% confidence."
Classification. Out of scope. The skill does not support predictive modeling. Recommend proper forecasting tools (Tableau's built-in Forecast feature via Analytics pane is close but should be flagged with its own uncertainty bounds).
Ambiguous questions
When the author's question matches two or more rules:
- Ask one clarifying question. Example: "Do you want a line at a specific value (W1), or do you want to see the projected trend (W3)?"
- Classify based on the answer.
- Do not pick for the author.
Non-standard questions
If the question doesn't match any rule and isn't obviously parameter-only, Monte Carlo, or predictive, state the skill cannot classify it in-scope and ask the author to rephrase or declare the mechanism manually.
Example response: "I cannot map this question to a mechanism in my catalog. Could you tell me more specifically what you want to see on the chart? For example: a line at a value, a band showing a range, a projection extending the trend, or an overlay comparing two scenarios."
Examples by classification
| Author's question | Classification | Reasoning |
|---|---|---|
| "Add a $1M target line" | W1 | Single value, "line" |
| "Shade the 25th-75th percentile" | W2 | Range, "shade" |
| "Where will revenue be in a year" | W3 | Projection, "where will" |
| "Who moves into the top-10 if they grow 20%" | W4 | Membership, "moves into" |
| "Overlay target on actuals" | W5 | Overlay, "overlay on" |
| "Create a calc for revenue +20%" | W6 | Baked-in field, "create a calc" |
| "Let the user choose growth 0-50%" | Parameter-only | Slider / user choice |
| "Simulate outcomes with random churn" | Out of scope | Monte Carlo |
| "Predict Q4 at 95% confidence" | Out of scope | Statistical prediction |
Audience note
The question classification is audience-agnostic. The author's audience shapes nothing in the W-mechanism choice. (Compare: stakeholder-translator and caption-and-tooltip-advisor are audience-aware; what-if-advisor is not.)