AgentSkillsCN

ecom-learning-loop

将创意素材、提示词、创意角度与效果标签持久化存储，以便日后基于现有数据生成新的变体（例如：“为我们的最佳创意角度再打造20个新版本”）。

SKILL.md
--- frontmatter
name: ecom-learning-loop
description: "Persist creatives + prompts + angles + performance tags so you can generate new variants from data (e.g., “make 20 new variants of our best angle”)."

Ecom Learning Loop (Creative Memory → Next Tests)

Goal: turn ad testing into a compounding system:

  1. •store what you made (angle, prompt, format, creator)
  2. •store what happened (spend, CTR, CPA, purchases)
  3. •generate the next tests from data, not vibes

Required output (create this file)

Maintain a registry file (recommended):

  • •outputs/meta-creative-registry.json

Each item should store:

  • •id, type (image/video/copy), source_path or URL
  • •tags: brand:*, product:*, angle:*, format:*, hook:*, stage:*
  • •performance: spend, impressions, ctr, cpa/cpp, roas, purchases, notes

Minimal tagging taxonomy

  • •brand:<slug>
  • •product:<slug>
  • •angle:<angle_id>
  • •format:<ugclook|product|proof|comparison|howto|video>
  • •hook:<hook_type>
  • •stage:<unaware|problem|solution|product|most>

Winner → 20 variants workflow

  1. •Pick the best creative for an angle (by CPA/ROAS) and document why it worked.
  2. •Generate a “subtle change” prompt pack: 10–30 variants preserving the core mechanism.
  3. •Launch and tag the variants with the same angle:<angle_id> so you can attribute learnings.

Operating rhythm

  • •Daily: log every launched creative + copy variant
  • •Daily: write back results (wins and losses)
  • •Weekly: export top angles + mechanisms and plan a new creative sprint