Source: raw/Claude_Cowork_Just_Changed_How_You_Do_Marketing.md; raw/CLAUDE_COWORK_FULL_COURSE_2+_Hours.md (skills, evals and marketing chapters); raw/How_to_Use_Claude_Cowork_Better_Than_99%_of_People_Full_Guide.md (newsletter and YouTube-packaging skill builds)
Creator: Ben AI · URL: https://www.youtube.com/watch?v=FAEKF-nr3D0 · Duration: 41:06 · Platform: YouTube · Uploaded: 2026-03-14 Also: CLAUDE COWORK FULL COURSE 2+ Hours (https://www.youtube.com/watch?v=2HyBA-wkWsA, 2:28:52, uploaded 2026-04-04) · How to Use Claude Cowork Better Than 99% of People (https://www.youtube.com/watch?v=HTu1OGWAn5w, 21:07, uploaded 2026-01-31)
Ben AI’s method for turning marketing work in Claude Cowork into reusable skills: do the task once with Claude, feed it a fixed set of reference documents, then ask it to save the process as a skill with rules that force those documents to be read. He walks the method through four skill types (copywriting, brand-aligned visuals, ideation, analytics) and adds a test-and-improve loop using the skill-creator’s built-in evals and A/B tests. The videos date from January to April 2026, before the September 2026 chat/Cowork merge, so treat the UI steps as historical (see the outdated-UI notes at the end) and the method as the durable part.
Key Takeaways
- Go into each task intending to build a skill. Ben’s framing is “onboarding an intelligent intern once”: the first run takes longer than doing the work by hand, because you are setting up connectors and writing instructions you will never have to repeat.
- Three ways to get a skill: customise one of Anthropic’s plugin skills (Customize → Browse plugins → Marketing → Install), download one from a marketplace (he names Smithery, SkillsMP and skillhub; the January video quotes Smithery at 14,500 skills), or build your own. He recommends building your own, and of the two build routes prefers “do the process once together with Claude, then ask it to build a skill” over describing the process up front. Converting an old Claude Project or custom GPT works too: paste its system prompt and knowledge files and ask for a skill.
- Reference documents decide output quality for copywriting. Ben keeps six or seven reusable context files and says they make skill output “10 times better” (his claim). His set: what we do (the business, used for calls to action), ICP, voice personality (tone attributes, the message readers should carry away, audience misconceptions, signature phrases), personal background (a four-to-five-page life story the skill can pull anecdotes from), channel strategy (newsletter, YouTube or LinkedIn strategy), writing framework, and examples of past output, which he calls the most important. LinkedIn skills add a hook template file; visual skills need a brand guidelines file.
- Build the reference files with Claude. Prompt: “I want you to create a YouTube strategy document that I can later give to you when I’m building skills for research, ideation, scripting, etc. Can you help me go through some questions that you want to ask me?” Or adapt someone else’s template: “This document is from someone else. I want to use this same format, but adapt it to me and my brand strategy. Please analyze the doc and ask me questions.” Budget 30-60 minutes per document.
- Ask for options at every human-in-the-loop step. Five angles, then three outlines, then the draft, then ten subject lines. Ben says choosing between options converges faster than iterating on a single output. Steps can be removed later as the skill gets reliable.
- Two phrasings he uses to pull output toward his voice: “almost mimic it” (the example documents) and “almost mimic the sentence structure, sentence length, and styling from the examples.” He also adds “base claims on what I said in the video, not make them up yourself” and a final self-check: “Before outputting the newsletter, please review and check the newsletter output across all the above references and guidelines.”
- Keep
SKILL.mdto the process; put everything else in reference files. Long context (a brand guide, MCP usage notes) goes in a separate reference file, and the skill names the step where it must be read. - Three rules to add when you ask for the skill: (1) reading the reference files is obligatory at the named steps (“it cannot skip over that”), because he finds skills otherwise skip them; (2) every correction made during the manual run becomes a rule in
SKILL.md; (3) a “progressive update” rule: when the user says not to do something, the skill adds that as a new rule. The course adds a fourth: when the user approves a final output, save it as a good example. - Test before reuse, one variable at a time. Ask Claude to test the skill and state what to optimise for, the pass criteria, and how to run the test. Then tell it to apply its suggested rule changes. Details under Building and testing the skills below.
- Commands chain skills; plugins carry them to teammates. Ben’s
/repurposecommand runs LinkedIn writer → newsletter writer → Excalidraw slides → GIF creator → infographic creator from one YouTube link. He shares a “creative” plugin so a teammate gets the skills, the commands and the Nano Banana connector in one import. He recommends building commands only after you have several working skills. - A CSV is enough of a database for a recurring skill. His newsletter-ideation skill logs every source it has checked so the next scheduled run skips them.
The newsletter-writer build, step by step
The worked example in both the marketing video and the course.
- Select a folder for the chat so outputs save to disk.
- Fetch the source. “Use the Apify MCP YouTube scraper to get the transcript.” (Apify setup: Settings → Connectors → Browse → Apify; paste the API token from Apify Settings; at mcp.apify.com tick the actors you want, then copy the part of the MCP server URL after the
=into the connector’s enabled-tools field; set the connector to Always allow.) - Angles. “Based on my ICP and newsletter strategy document, suggest me five potential angles for this newsletter based on this YouTube video.”
- Outline. “Give me three suggestions on high-level outline of sections… But before doing that, please read my newsletter example doc.”
- Draft, with the voice instructions above, plus: “closely follow the voice personality guide”, “read my personal background doc… potentially some stories you could use, if appropriate”, “reread my ICP document and my newsletter strategy document”, and “read the what we do doc” for the call to action.
- Subject lines. Ten options from the examples file. Ben corrects the first set (“between three to eight words”) so the correction lands in the skill as a rule.
- Save as a skill. His prompt names the skill (“Ben’s newsletter writer”), its trigger (“every time that I mention that I want to write a newsletter”), says to follow the exact process above and include every reference file from the chat, and adds a new first step: ask whether the source is an idea, an article or a YouTube video, and scrape articles with the Apify web scraper. Then the three rules above.
- Test, update, and replace the saved skill (“Copy to your skills” → upload and replace makes a new version).
The resulting SKILL.md steps: identify the source → read reference files → suggest five angles → read more reference files → suggest three outlines → write the newsletter → suggest ten subject lines. The January video shows an earlier, shorter version (subject lines → hook → newsletter) that fetched transcripts by driving a transcript website through the browser, which Ben says an MCP or API does better.
Brand-aligned visuals
- Same method, different stops. For his LinkedIn infographic skill: “give me five suggestions on which part of this post would make the strongest visual” → five ways to visualise the chosen part → three ways to structure the copy → generate.
- Images come from Nano Banana, not Claude. Ben says Claude cannot generate PNGs and uses Google’s Nano Banana through an MCP, which he calls “by far the best” image model (his opinion, March 2026). It needs a Gemini API key from Google AI Studio; there are free credits, then billing is required.
- Brand guidelines file. He fills a brand-guideline template with Claude (give it your website or existing assets to start), then asks the skill to turn the brand guide and a “how to prompt the Nano Banana MCP” guide into separate reference files.
- Two extra first steps in the saved skill: ask for the content or idea, and ask which platform it is for so the image gets the right format.
- Iteration count. He says he updated the infographic skill five to ten times before most outputs were usable, and that the v2 skill carries many more reference files than v1. The course shows the later version using multi-select question boxes so several variations generate at once, and a “keep all / finalise” step that saves approved outputs as examples.
- No connector needed for most formats. He says Claude produces PowerPoint and Google Slides files, HTML dashboards, and Excalidraw JSON (paste into Excalidraw for branded slides) without external tools; only raster images need the Nano Banana connector. Other visual skills he shows: client case studies from a call transcript, GIFs, and branded Excalidraw decks from transcripts or outlines.
Ideation and research on a schedule
- Point the chat at a context folder instead of attaching files each time (his “Ben AI context copy” folder holds the ICP and newsletter strategy).
- Prompt shape: give five sources he reads (sites such as a16z and Anthropic), ask Claude to find content published in the last 24 hours, qualify each piece as relevant or irrelevant against the ICP and newsletter strategy, and output “a well-structured HTML report that I can open in my Chrome.” The run found seven newsletter ideas with suggested stats and infographics.
- Dedup database. Before saving the skill: “I plan on running this every 48 hours… please create a CSV that consists of all the qualified sources and unqualified sources with the date,” with columns for source name, URL, date, execution time and qualified/unqualified, and an instruction to check it first on every run. The CSV is stored as a reference file in the skill.
- Schedule it by telling Claude (“schedule this skill every day at 8:00 a.m.”) or through the Scheduled sidebar (name, prompt “Use the newsletter ideation skill to ideate for my newsletter”, frequency, folder, model). See Scheduled Tasks for the Cowork scheduling mechanics.
- Competitor tracking uses the same pattern with Apify scrapers for LinkedIn and YouTube posts.
Analytics and SEO
- YouTube analytics. He exported a YouTube Studio CSV, asked which analyses would be useful, then had Claude build an HTML dashboard with Anthropic’s built-in frontend-design skill. Because the skill can also pull transcripts through Apify, it compared performance with what he says in his intros: top videos by views, like, comment and share rate; subscriber conversion and CTR; upload frequency; best video length; and the finding that his best hooks contain numbers or statistics and that how-to tutorials perform best. He saves it as a monthly skill.
- Other analytics he lists for the same treatment: SEO reporting and Meta ad reporting, on a weekly or monthly schedule.
- SEO audit skill covering legal compliance, performance, E-E-A-T, content and GEO readiness, returning critical issues, high and medium priorities, and quick wins. He pairs it with an SEO-optimiser skill and the Webflow MCP and says the pair “can almost autonomously” keep his site optimised. He says programmatic SEO and SEO-optimising skills work better in Claude Code than in Cowork.
Building and testing the skills
From the course’s skills chapter, which generalises the newsletter build.
Skill prompt framework (Ben says it is “not hard science”):
- Name and trigger (“should be triggered anytime a user mentions he wants to generate or create an infographic”); Claude writes the description from this.
- Goal, one or two sentences.
- Connectors, APIs or MCPs, including which table or page to use.
- Process, step by step: what happens, where the human steps in and in what form (checkboxes, single select, open field), which reference file each step reads, and the expected output.
- Rules: predicted failure modes, obligatory reference reads, multiple options at human steps, progressive updates, and saving approved outputs as examples.
Reference file types a skill can carry: text files (examples, style guides, ICP, background, and an “MCP instructions” file on how to use a tool for this specific process), assets (images, decks, video), and scripts (his infographic skill includes a script that calls the Nano Banana API).
Evals with the skill-creator. Ben calls the updated skill-creator “skills 2.0”: after building a skill, Claude offers to test it, spins up parallel sub-agents, grades each run and produces a report with per-run feedback fields you can copy back into the chat. His guidance:
- An unspecified test only checks that the skill runs; Claude invents its own criteria (in his run: a P.S. pitch, a sign-off, word count, a .docx file).
- To optimise, state one thing to optimise for, the criteria, and how to run it. His example: optimise for matching his voice; criteria = closeness to the examples, em dashes, length, personal stories from the background file; run five variations on one YouTube video. Result: style match failed two of five, word count and personal stories one of five each.
- Then: “Please update the skill according to the rules you suggested,” re-save, and repeat.
A/B tests compare skill versions; he runs them only once a skill already works, and notes Anthropic suggests rerunning them when a new model ships, since a skill can stop adding value. Two of his tests:
- Speed. “Use the skill-creator skill to run an A/B test on the YouTube to newsletter skill so that we optimize the skill for speed. It can’t affect the step-by-step process.” Claude built a leaner variant: 77,000 tokens and 160 seconds against the original’s 93,000 tokens and 204 seconds (his figures). The lean variant failed transcript extraction and word count; Claude attributed the transcript failure to the scraper behaving differently per agent, a rerun passed, and he kept the lean version.
- Context. All eight reference files versus the same set without the voice-personality file, to see whether that file helps. He notes the verdict on copy quality is subjective and needs a human read.
For longer optimisation he runs an auto-research loop (hypothesis → test → keep or discard) covered in Ben’s feature primer and Karpathy autoresearch for cold outbound.
Outdated UI in these videos
The method still applies; these specifics have changed or may have changed since the January-April 2026 recordings.
- Ben describes Cowork as a second desktop tab beside Chat, desktop-only and unavailable on the Free plan. Cowork reached web and mobile in July 2026, and chat and Cowork merged into one Claude in September 2026 (rolling to Pro, Max and Team through October). See Claude Cowork.
- His scheduled tasks run only while the desktop app is open. Since July 2026 scheduled tasks in cloud projects run with no device online, and since October they attach to projects. See Claude Cowork.
- “Copy to your skills”, the Settings → Capabilities → Example skills list and the model picker offering “Opus 4 or Sonnet” are early-2026 UI.
Try It
- Write your “what we do”, ICP and voice-personality files first, using the question-asking prompt above; collect five to ten examples of your best past output in one document.
- Pick one weekly copy task. Run it once in a folder-backed chat, asking for options at each step and correcting anything off.
- Save it with the build prompt: name, trigger, exact process, all reference files, obligatory reads, corrections as rules, progressive updates.
- Run one eval with a single optimisation target and three to five variations on the same input; apply the suggested rules.
- Once two or three skills work, chain them in a command and share them as a plugin.
- For any recurring research skill, add a CSV log of what it has already processed before you schedule it.
Open Questions
- Ben’s “10 times better” claim for reference files and his A/B token and time figures are single runs on his own skills; no wider measurement is offered.
- Which reference files matter most? He tested removing the voice-personality file but did not report a conclusion.
- Nano Banana pricing and quality are his March 2026 view; image models and Claude’s own image abilities may have changed since.
- Skill self-updating. The progressive-update rule asks a skill to rewrite its own
SKILL.md; the videos do not show how this behaves for a skill shared through a plugin or an organisation, where the user may not own the file.
Related
- Cowork Sales Skills and Bulk Sub-Agents (Ben AI) — the same method applied to prospecting, follow-up and sales analytics
- Claude Cowork for Marketing — Anthropic’s eight marketing use cases
- Every Claude Cowork Feature Explained (Ben) — Ben’s later feature primer, with his other Cowork videos mapped
- The Complete Guide to Building Skills for Claude — Anthropic’s skill-creator and eval background
- Skill Authoring Best Practices
- Cowork + Apify Scraping Recipe
- Nano Banana
- Cowork Plugins