Performance Marketing Eval — AI skill for Claude Code

performance-marketing-eval

A benchmark by **GoMarble**.

15 tasks testing whether AI agents can analyze marketing performance and safely manage accounts across Meta, Google, TikTok, and Shopify.

Tasks run locally against fixed data and simulated accounts.

What it measures

Accurate numbers, complete data retrieval, useful explanations, safe changes, and recovery from failures.

Case	What it tests
001	Purchases and cost per purchase
002	Duplicate search queries and wasted spend
003	Campaign creation without increasing total budget
004	Bid-change timing and names matching live settings
005	Creative cleanup and reversing failed budget increases
006	Retrying without creating duplicate ads
007	Restoring bid strategy while preserving the CPC cap
008	Counting new ads and identifying their creators
009	Verifying a merged ad set before retiring originals
010	Budget pacing after corrected finance figures
011	CTR improvement versus changes in traffic mix
012	Funnel losses by device and landing page
013	Daily budgets without double-counting shared budgets
014	Conversion traceability to keywords and searches
015	Reconciling complete and partial financial reports

Full prompts, IDs, and rubrics: [cases.jsonl](data/cases.jsonl).

Directory structure

data/     Tasks, datasets, tools, and scoring rules
src/      Runners, simulated tools, and grader
results/  Baseline scores and full traces
runs/     Your local runs and grades (gitignored)

Run locally

1. Install Bun 1.2+, Node.js 20+, and Claude Code or Codex.
2. Configure your CLI login or API key.
3. From this repo, run either example:

bun run eval:claude --model claude-sonnet-5 --effort high --out runs/claude
bun run eval:codex --model gpt-5.6-terra --effort high --out runs/codex

Use the full model ID supported by your CLI account. Add `--tier sanity` for two quick cases or `--cas