Recursive Self-Improvement Loops for Marketing with Claude

Summary

A technique for using Claude where instead of prompting once and shipping, you build a recursive loop: generate output, evaluate against a scoring criteria, diagnose weaknesses, rewrite, re-evaluate, and repeat until it passes. Example: an image ad skill scores every concept against 10 criteria (thumb-stop power, curiosity gap, emotional trigger, persona recognition) with a 9/10 minimum threshold. If it fails, it explains why, rewrites, and re-scores. The pattern applies to email sequences (stress-test subject lines, hooks, CTAs), video hooks (retention psychology for first 3 seconds), positioning (adversarial testing against competitors), and SEO content (search intent, competitive gap analysis). Build it once, save as a reusable skill.

Why This Was Bookmarked

“another article on how to use Claude for marketing”

The recursive self-improvement pattern is the marketing-specific version of the adversarial review pattern from the Anthropic skills article. Directly applicable to 01-projects/newsletter/index content quality.

Key Ideas

• Recursive loop: generate > evaluate > diagnose > improve > repeat
• Scoring criteria with thresholds: “thumb-stop power: 9/10 minimum” not “make it good”
• Adversarial pressure: give Claude a persona that attacks the output (skeptical customer, competitor, distracted buyer)
• Domain-specific applications: email drips, video hooks, positioning, SEO content
• Save as reusable skill: don’t prompt from scratch each time — build the loop into a skill file
• Build process: pick a repeatable task > write your evaluation criteria > set pass/fail thresholds > build the loop > add adversarial pressure > save as skill

Connections

This is the marketing-domain implementation of the verification skills pattern from 06-reference/2026-04-04-anthropic-skills-internally. Anthropic says verification skills are highest-ROI; this shows what that looks like for marketing content.

The recursive loop pattern is compound engineering applied to content: each iteration makes the output better, and the skill itself captures the criteria for reuse. See 06-reference/concepts/compounding-knowledge and 06-reference/2026-04-04-compound-engineering.

The adversarial pressure technique connects to 06-reference/concepts/skills-as-building-blocks — a well-designed skill includes its own quality gates.

For 01-projects/newsletter/index: we could build a recursive skill for newsletter issue quality that scores against our specific criteria (actionable insight, data-backed, voice consistency, reader value).

Open Questions

• What scoring criteria would we use for newsletter content quality?
• Can we combine this with the vibe marketing system’s research > skills > build pipeline?
• Should we build a generic “recursive quality loop” skill that wraps any content-producing skill?