# Marketing experiment design

Original prompt workbook by Jon Bishop. CC BY 4.0. Model execution has not been tested.

## Intended output

A proposed design with assumptions and missing statistical inputs

## Input worksheet

Decision, assignment unit, primary metric definition, baseline data, eligibility, constraints and practical effect threshold

Fill in:
- Decision and owner:
- Source ID, URL/file, date and exact supporting observation:
- Relevant population, version or edition:
- Approved constraints and sensitive fields to remove:
- Missing evidence:

Use sanitized inputs. Follow your approved model provider and data handling rules. No external action is part of these prompts.

## Prompt 1

```text
Inspect the supplied inputs for marketing experiment design. Identify supported facts with their source IDs, contradictions and missing information. Do not infer missing product capabilities, numbers or customer evidence. Return an evidence ledger and blocking questions.

SUPPLIED INPUTS / PREVIOUS ARTIFACT:
[paste sanitized material here]
```

## Prompt 2

```text
Using only the evidence ledger and the original inputs, prepare a proposed design with assumptions and missing statistical inputs. Label assumptions and unknowns. Keep dates, versions, cohort and edition boundaries attached to claims. Do not execute external actions.

SUPPLIED INPUTS / PREVIOUS ARTIFACT:
[paste sanitized material here]
```

## Prompt 3

```text
Review the draft against this criterion: Assignment, denominator, contamination and stopping rules are explicit; qualified review is needed for inferential claims. List failed checks, unsupported claims and corrections. Return a corrected draft plus unresolved questions; do not claim checks were performed when the inputs do not contain their results.

SUPPLIED INPUTS / PREVIOUS ARTIFACT:
[paste sanitized material here]
```

## Worked example — fictional, manually authored

Input: Fictional: ten weekly eligible trials make a quick multi-variant test unlikely to be informative.

Expected bounded output: Recommend checking instrumentation and traffic feasibility before choosing sample size; do not invent a statistically adequate duration.

This illustrates the review standard; it is not a recorded model run.

## Human review

Assignment, denominator, contamination and stopping rules are explicit; qualified review is needed for inferential claims.

Verify factual claims against the original evidence, check calculations or commands independently and record unresolved issues. Save the inputs, prompt version, model version and corrections if you run this workbook. Publication, sending, account changes and spending are separate decisions.