Skip to contents

Renders the seven versioned prompt templates shipped in inst/prompts/, one per liminal dimension, with the business ask and the optional situation interpolated as data. The templates embed a fixed anchor rubric and a fixed direction statement, so scores are comparable across business cases (see the package specification).

Usage

liminal_prompts(ask, situation = NULL)

Arguments

ask

A single non-empty character string: the business ask, verbatim, in whatever language it was written.

situation

NULL or a single character string with context, data state and known constraints. When NULL, the prompts state that no additional situation was supplied.

Value

A named list of seven character strings, one per dimension, in canonical dimension order.

Details

The rendered prompts are plain text; scoring them against a model is the job of assess_liminal_space().

Examples

prompts <- liminal2::liminal_prompts(
  ask = "Build us a churn model so we can prioritise retention spend."
)
names(prompts)
#> [1] "shared_meaning"            "stable_ground"            
#> [3] "convergent_answers"        "knowable_from_observables"
#> [5] "judgeable"                 "action_identifiable"      
#> [7] "tradeoff_agreed"          
cat(prompts$shared_meaning)
#> You are scoring one dimension of the liminal space between a business
#> ask and a data-based solution. The liminal space is the gap between
#> what the business asked for and what a data-based solution can
#> deliver. You score only the dimension below, not the overall gap.
#> 
#> ## Dimension
#> 
#> Shared meaning - rival operationalisations of the target agree.
#> 
#> Question a business owner can answer: When we say "churned" or
#> "risky", do we all mean the same customers?
#> 
#> Question a data scientist can answer: Do rival label definitions
#> agree on who is positive?
#> 
#> Evidence: label agreement across rival definitions (kappa or rank).
#> 
#> ## Scale
#> 
#> Return an integer from 0 to 10. 0 means the dimension poses no
#> obstacle (the smallest possible liminal space). 10 means the dimension
#> is maximally obstructed (the largest possible liminal space).
#> 
#> ## Anchor cases
#> 
#> Score 2 (small space): One authoritative definition exists in a data
#> dictionary and every downstream use follows it.
#> 
#> Score 5 (moderate space): Two competing definitions are in use; the
#> choice moves a modest share of cases and is undebated.
#> 
#> Score 8 (large space): Three or more rival definitions are live and
#> the choice flips a material share of cases; no owner can arbitrate.
#> 
#> ## Calibration
#> 
#> Match the structural condition in the anchor, not the amount of
#> missing information. The 8 anchor is a specific structural case, not
#> a default for an ambiguous ask. If the input does not meet that
#> condition, score below 8. When the evidence falls between 5 and 8,
#> use 6 or 7; do not jump to 7 or 8 merely because something is
#> unspecified. Score this dimension on its own evidence: a defect that
#> belongs to another dimension must not raise this score.
#> 
#> ## What counts
#> 
#> Count only rival definitions of the target label that change who is
#> classified positive. Two definitions are rival only when they would
#> label a different set of cases positive.
#> 
#> If the ask requests descriptive quantities rather than a positive
#> class (counts, hours, rates, indicators), there is no label to
#> operationalise: score 1-2 unless the input also supplies rival
#> definitions of those quantities that would produce materially
#> different numbers. An undefined normative word such as "well" or
#> "execution" is not by itself a rival label definition when the ask
#> names the quantities to compute.
#> 
#> If the ask names a single target but the situation describes several
#> positive classes with different label paths (for example, several
#> fraud types under one "fraud" label), treat them as rival
#> operationalisations only when choosing among them would change which
#> cases are positive. If no owner can arbitrate that choice, score 6-8.
#> 
#> Do not count as shared meaning:
#> - the performance horizon or observation window;
#> - the decision the output will drive;
#> - the required output granularity;
#> - the absence of a baseline or of a definition of "normal";
#> - which features are available or how they are joined.
#> 
#> ## How to score
#> 
#> 1. Quote the definition of the target that the input states or
#>    implies. If the ask is descriptive, quote the quantity requested.
#> 2. Quote any rival definition. If you cannot quote a second
#>    definition that changes who is positive, score 1-3 for a descriptive
#>    target, or below 5 for a label target.
#> 3. If the input explicitly says the label is settled or externally
#>    defined, do not score above 4 on this dimension.
#> 4. Otherwise match the quoted definitions to the anchors and
#>    interpolate between them.
#> 5. Reserve 0-1 for a case where every downstream use follows one
#>    written definition, and 9-10 for a case where rival definitions
#>    are irreconcilable and no one can arbitrate.
#> 
#> ## Input
#> 
#> Business ask:
#> 
#> ```
#> Build us a churn model so we can prioritise retention spend.
#> ```
#> 
#> Situation and context:
#> 
#> ```
#> No additional situation was supplied.
#> ```
#> 
#> ## Output
#> 
#> Return only the integer. No explanation, no units, no decimals.