Source: raw/Claude_+_New_Al_Marketing_Agent_Does_80%_of_My_Work_Letaido_Agent.md (https://www.youtube.com/watch?v=vXMSxajeL7M), YouTube auto-captions fetched 2026-10-30. Creator: Grace Leung (youtube.com/@graceleungyl) Platform: YouTube Duration: 16:43 Published: 2026-06-30 Skill level: Beginner (hosted product; no setup)

Sponsored product demo

“For this video, I have partnered with Ahrefs.” The whole demo runs inside Ahrefs’ own product, so product claims here are the sponsor’s showcase, not an independent test. The method (skills → agent roles → tools) is product-neutral and is the reusable part. The raw file’s title reads “Letaido Agent”; the spoken name throughout is “Ahrefs Agent A”, which this article uses.

Agent A is Ahrefs’ hosted AI marketing agent, built on Ahrefs data, with ready-made skills, apps and connectors and a shared team workspace. Grace Leung uses it to show three levels of getting AI to do marketing work: run skills, define agent roles, then turn workflows into tools. She ends with a straight build-versus-buy answer: Claude plus the Ahrefs MCP if you are technical or solo, a hosted product like Agent A if you are a team that doesn’t want to maintain infrastructure.

Key Takeaways

  • The hard part of agents is upkeep. “The hardest part is not building it, it’s maintaining it. New tools, changing models, broken connections, constant babysitting.”
  • Three levels, platform-independent:
    1. Leverage agent skills. Run proven workflows on command; “your quickest, most immediate ROI.”
    2. Define agent roles. “Skills … on their own, they’re just actions. They don’t know when to do what.” A role decides which skills to use.
    3. Build tools. Package a workflow into an app that anyone on the team can run, so “your best workflows stop living inside the chat window.”
  • Map your work before touching skills. “Most people run AI as one endless chat thread.” Give the agent context about your day-to-day work and ask it to organise it into a skill library grouped by marketing discipline.
  • Six elements of a reliable agent role: its role, the input it takes, the skills or tools it uses, the workflow it follows, the output you want, and the guardrails it stays inside.
  • Accurate data is the selling point she keeps returning to. “Every AI models can build this kind of dashboard, but that doesn’t mean the analysis is accurate.” Her case for Agent A rests on Ahrefs’ data behind the skills, not on the agent itself.
  • Model choice changes the bill. Agent A’s model picker includes Claude. Her AI-search gap analysis with Claude Opus “feels a bit costly”; she suggests a lighter model “like Sonnet or even ChatGPT”, which she found “much more affordable” with results that were “pretty good as well”.
  • Build vs buy, in her words. “Why not just use Ahrefs MCP or other data API inside Claude? … Yes, you can.” Build it yourself if you are technical, work solo or want full control of the stack and model. Choose a hosted agent if you are on a team, don’t want to babysit infrastructure as tools and models change, and want data, skills and workflows ready on day one.

What she demos (demo brand: Caraway cookware)

  • Level 1, pre-built skill. An AI-search visibility gap analysis: enter brand, domain and the AI platforms to check. It returned mention gaps and content opportunities by priority, with the target questions to answer, plus a dashboard in the Team Console panel (search by question, sort by priority, competitor or platform). Follow-ups: propose three topic ideas to close the gap, and set the analysis to run monthly.
  • Level 1, chained custom skills. She uploaded her own brand-strategy-deck skill to the workspace knowledge base (which Agent A calls the wiki). She then asked for a full AI-search growth strategy (visibility, mention gaps, citation performance) packaged as a deck. Result: 14 branded slides using Ahrefs data. She still brainstorms direction and narrative with AI before finalising a deck.
  • Level 2, a content-strategist role. The role chooses which topics the brand should own, has default skills and is saved as a playbook in the wiki. She updated Agent A’s memory so the playbook loads whenever she tags the role. Run on the gap analysis, it produced a 90-day topic-cluster dashboard ranked by brand fit and effort; “non-toxic cookware” and “best non-stick cookware” came out on top.
  • Level 2, a blog-writer role. It picked “now”-priority items from the opportunity board, did SERP and internal-link research with Ahrefs data and its AI content helper, and saved a WordPress draft through the WordPress connector. The draft followed her blog-writing skill: key-takeaway bullets, question headings, tables, images, internal links and an FAQ. She calls the connector list “quite limited” for now and says to “always review it with your own judgment before actually posting.”
  • Level 3, apps. A pre-built content keyword research app took a niche and seed keywords and returned clusters by traffic potential, a competitor view, keywords tagged with pillar, intent and recommended format, a hub-and-spoke map and a PDF export.
  • Level 3, a custom tool. She asked Agent A to wrap a community-content research skill into a tool: research forums for real questions, propose angles, wait for her approval, then generate carousel copy and visuals. It took “about 10 minutes” to build, and the visuals download in one click. Teammates in the shared workspace can use the tool, and a team whiteboard organises tools and client reports.

Try It

The three levels work in Claude without Agent A; this mapping is the wiki’s, not the video’s.

  1. Skills. Ask Claude to turn a description of your week into a skill library grouped by discipline; build the first two (see her Claude Code build).
  2. Roles. Write one agent file using the six elements (role, input, skills/tools, workflow, output, guardrails) and route to it from CLAUDE.md.
  3. Tools. Turn the workflow you run most into an interactive artifact or small app that teammates can use without opening the chat.

Open Questions

  • Agent A’s pricing and credit costs aren’t given; she shows a credits counter but no figures.
  • The full list of connectors, apps and models in Agent A isn’t shown, and she says the connector list is “quite limited” as of June 2026.
  • The video doesn’t say whether Agent A runs Anthropic models through the API (and which versions), beyond “Claude Opus” and “Sonnet” in the picker.
  • No independent check of the AI-search gap data quality; the accuracy claim is the sponsor’s pitch.
  • The raw file’s title says “Letaido Agent”. Whether the published video title uses a different product spelling couldn’t be confirmed from the transcript.