LinkedIn Ads Creative Sprint — 20 Ad Variants Across 5 AI Image Models with a Scored Rubric
Run a single-session creative sprint across five AI image models (plus an in-house HTML lane) that produces finished ads, a scored rubric for each, and a shortlist defensible against the brief.
What was getting in the way.
AI image generation has become a tool-shopping problem: every model has different strengths at text fidelity, concept relevance, photorealism, and typography, and nothing picks the right one for a given ad concept except head-to-head testing. Running that test by hand means hours of prompting across tools, inconsistent judgment, and no scorecard anyone else can audit.
Context
Universal — applies to any AI agent deployment running structured creative sprints where head-to-head scoring across tools matters.
How the work runs.
- 01
Trigger
a request lands for a batch of ad creatives across several concept angles — in this case, 4 angles × 5 lanes = 20 ads.
- 02
Fix the rubric before the first render
a 6-row scoring rubric covers thumb-stop, 3-second message, brand fidelity, audience/angle fit, craft & platform fit, and visual-concept relevance; double-weighted rows the brief names.
- 03
Run every lane with the same brief
an in-house HTML pipeline plus four generative AI image models each produce four concepts; prompt law bars the stale brand marks, requires concept-first visuals, and forbids micro-text invention on photoreal screens.
- 04
Score at the full-size read, not the thumbnail
every ad judged at a feed-scale copy of itself before scoring, so legibility defects can't hide behind the hero crop.
- 05
Publish the scorecard with the sheet
the one-sheet scored PDF is a required deliverable — captions carry per-ad score, lane and classification, winners green.
- 06
Measure spend exactly
every generative lane's credit cost is simulated before spend and reconciled against the actual draw after; nothing is estimated.
- 07
Outcome
a reviewable creative sprint with a shortlist, a scorecard, and a reproducible method — ready to re-run for round two.
Evidence from the workflow.
Each system has a role.
Action (lanes)
Multi-model AI image generation
Action (control lane)
In-house HTML rendering pipeline
Record of truth
Scoring rubric + results-PDF generator
Signal (budget guard)
Dry-run cost estimator
Why this is Coworker.
Takes assignments and reports back. You hand it work, it prepares and returns a result.
- 01
No ad ships to a platform from this workflow — the AI's output is a scored shortlist a human picks from. Prompt-law bars stale brand marks and forbids micro-text invention on photoreal screens. The scorecard is reviewable before any media decision is made.
Impact / Outcomes
20 finished ad variants delivered the same session across 5 generation lanes against a shared 6-row rubric.
Round 1 batch average 28.5 vs round 2 batch average 32.2 under the same rubric; relevance sub-score moved from 2.85 to 4.70 out of 5 as the rubric was updated and winners re-chosen on evidence.
Every lane's credit cost simulated before spend and reconciled to the exact draw after; zero estimation in the spend report.
Scorecard is a required deliverable, not an afterthought — the one-sheet PDF ships with every contest.
Find where a workflow like this fits.
Start with the systems, work, constraints, and authority already present in your operation.