Eric Siu grew X impressions to 3.1 million using a reusable skill

A marketing team writes the same ChatGPT prompt from scratch every week, and when the person who wrote it leaves, the knowledge leaves with them. Eric Siu, the founder of the agency Single Grain, published his fix on Moz: save every repeated process as a skill, score it with an eval, and chain the skills into loops that keep improving.

Siu says one skill, written for long-form posts on X, took his own account from 180,000 impressions in January to about 3.1 million in June. The result keeps compounding because the team keeps rewriting the skill, Siu says. The X skill is now one of about 2,600 open ones in his GitHub repository of AI marketing skills.

What a skill is

A skill is a Markdown file that an AI agent reads and runs the same way every time. Anthropic released Agent Skills as an open format, Siu writes, and OpenAI then added support in ChatGPT and Codex. Siu's X long-form skill takes a topic and a contrarian angle, drafts a post with a hook and ASCII diagrams, then runs the draft through a second skill called a humanizer that rewrites anything that reads as AI.

Siu says the team keeps every skill in one repository and lets anyone improve them. He rates his own X skill a five out of ten; one of his marketers has taken a version of it to an eight or nine. Siu writes that new hires get a bundled pack of 20 skills on day one, which is how the agency skips a six-month onboarding ramp.

Evals score the output

An eval is a definition of done, Siu writes. The humanizer on the X skill is a small eval: every draft starts at 100 and loses points for each AI tell, and the piece ships only at 90 or higher. If a draft falls below, the skill rewrites it and scores it again.

Siu writes that one of his customers moved from about 5 percent to 20 percent share of voice in AI search once a loop ran this way. He notes that the share-of-voice metric is widely used but imperfect. The number he puts behind the stakes is from Pew: Google users clicked a traditional link in 8 percent of visits when an AI summary appeared, against 15 percent when one did not.

Loops that learn

A loop is a chain of skills with a target state, Siu writes. His sponsorship loop replies to every podcast and YouTube sponsorship request with a flat $25,000 quote, logs who it responded to, and shows the fine-tunes it made each day. Siu writes that the loop learned on its own to send a media kit after sponsors kept asking for one.

A shared workspace holds everything. Siu says Single Grain built its "brain" inside Slack so no one has to log into a new tool, and that the agency uses OpenRouter to send each step to the right model instead of running everything on the most expensive one.

Teams that treat AI as a prompt they type each time will keep paying for the same lesson. Teams that write the lesson down once, score the output, and let a loop iterate will take the saving next week. The method is open: Siu made the Single Grain library free at skillsdojo.com.

Source

Operationalizing AI: Proven Workflows That Generate Measurable Revenue, Eric Siu, Moz.