Prompt Experiment Readout for Media

The Prompt

You are an experimentation analyst. Turn media prompt test results into a decision-ready readout that says what to keep, what to drop, and what to test next.

EXPERIMENT GOAL:
{{experiment_goal}}

VARIANTS TESTED:
{{variants_tested}}

EVALUATION DATA:
{{evaluation_data}}

OBSERVED FAILURES:
{{observed_failures}}

DECISION DEADLINE:
{{decision_deadline}}

NEXT ITERATION CONSTRAINTS:
{{next_iteration_constraints}}

Return exactly:
1) Experiment framing check
   - what the test was really trying to learn
   - what weakens the readout
2) Executive summary
   - what worked
   - what did not
   - what to do next
3) Variant comparison
   - variant
   - performance summary
   - confidence
   - keep, revise, or retire
4) Failure pattern analysis
   - recurring failure
   - likely root cause
   - where in the workflow it appears
5) Decision recommendation
   - scale
   - revise
   - stop
   - why
6) Next experiment plan
   - max 3 focused tests
   - what each test isolates
   - success signal
7) Sparse-data fallback
   - what can still be concluded
   - what more instrumentation is needed

Rules:
- Keep conclusions tied to evidence.
- Distinguish strong signals from weak or ambiguous signals.
- Stay tool-agnostic and avoid vendor-specific assumptions in the analysis.
- If the experiment mixed too many variables, say that the readout is not clean.

When to Use

Use this after running prompt trials for generated images, videos, or audio outputs. It helps teams avoid endless tweaking and move to clearer next-step decisions.

It is especially useful in current reference-heavy media workflows where generation, editing, and review can all hide the real cause of success or failure unless someone writes the readout carefully.

Variables

Variable	Description	Example
experiment_goal	The hypothesis or decision the test aimed to inform	”Does reference-image lane improve approval rate without slowing production too much?”
variants_tested	Prompt variants and setup differences	”Base prompt, reference-driven prompt, shorter-shot prompt”
evaluation_data	Reviewer notes, scores, acceptance rates, timing, cost, or engagement data	”Pass rate, revision count, render time, reviewer comments”
observed_failures	Recurring technical, creative, or compliance issues	”Text rendering drift, subject inconsistency, awkward dubbing timing”
decision_deadline	When the team needs a decision	”Before next week’s production sprint”
next_iteration_constraints	Budget, timeline, staffing, or tool limits	”Can only run 2 more test rounds this month”

Tips & Variations

• Ask for a minimum viable next test when time is limited and the team only gets one more round.
• Include confidence labels to prevent the readout from sounding more certain than the data allows.
• For small teams, request a one-page version with the top three actions only.
• If the workflow spans generation and human editing, ask the model to separate model failures from operator or review-process failures.
• If the data quality is poor, ask for instrumentation recommendations before running the next round.

Example Output

Decision: revise, do not scale yet. The reference-driven lane improved composition quality, but the experiment changed both shot length and prompt structure, so the causal read is still messy.

Next test: keep the reference pack fixed and vary only prompt density across two short scenes.