The skill-creator Skill: Build Real Workflow Skills in Plain English

The skills overview showed you the dad-joke demo — useful to understand the mechanics, useless for actual work. This module gets real. We’ll build three production-grade workflow skills using skill-creator, the Anthropic-built skill whose only purpose is to make other skills. After this you’ll have the template for every skill you’ll ever build.

Why skill-creator exists

You could write a skill by hand. It’s a markdown file with metadata — not hard. But skills have conventions (how the metadata is structured, how examples should look, how to phrase behavior notes) that matter for whether Claude actually invokes the skill later. skill-creator knows the conventions. You describe the workflow; it writes the file.

Two rules matter more than anything else:

1. Say “use the skill-creator skill” in the prompt. Early on, every time. Without the phrase, Claude often reasons about the workflow with general intelligence and writes a weaker skill.
2. Watch the sidebar. If you don’t see “reading the skill-creator skill” in the sidebar, it isn’t being invoked. Stop, fix the prompt, and try again.

Before you start: turn it on

Customize (briefcase icon) → Skills. Under Example skills from Anthropic, find skill-creator and confirm the toggle is on. Without it, none of this works.

Skill 1: The research analyzer

The most common knowledge-worker ask: “I have a 40-page research report; extract the stuff I actually care about in a consistent format.”

Prompt:

Create a new skill using the skill-creator skill.

The skill helps me review market research reports.

When I give you a research report, extract:

1. A summary of the key takeaways from the perspective
   of a long-only investor.
2. Any predictions on stock prices, as bullets, with
   the time horizon for each.
3. Every company covered, with bull/bear/neutral
   sentiment, a short explanation, and any price or
   performance target.
4. Any extreme negative scenarios the report
   describes, with associated market predictions.
5. Data points worth tracking going forward.

I'm attaching a sample report so you can see the
kind of input this skill will receive.

Attach a real report. The attached example is what makes this skill good — skill-creator uses it to calibrate tone and depth.

What happens. skill-creator writes a file named something like market-research-reviewer.md with five sections matching your spec, plus processing guidelines Claude uses when invoking it. Click Copy to your skills.

Verify. Start a new chat. Upload a different research report. Prompt:

Review this research report.

You should see “reading the market-research-reviewer skill” in the sidebar, then five sections appear. If it’s wrong or shallow, edit the skill with natural language:

Edit this skill — add a table at the top showing
predictions for the major stock indices before the
company-level section.

skill-creator revises. Click Copy to your skills → confirm Replace. The old version is still live until you replace.

Skill 2: A writing-voice style guide

The second-most-useful skill for most professionals. Teach Claude how you write.

Upload three to five pieces of writing you want to mimic — your own published pieces, or a colleague’s if you’re working on their behalf, or a canonical author whose voice your firm uses. Three is enough; five is plenty.

Prompt:

Here are five memos I want to use as a style
reference. Create a skill using the skill-creator
skill that captures this writing voice.

Pay attention to:
- Word choice and preferred verbiage
- Sentence construction and rhythm
- Paragraph length and flow
- Use of metaphors and analogies
- Tone (formal/informal, hedged/direct)
- Structural patterns (hooks, sign-offs, transitions)

Include 2-3 short examples from the memos inside
the skill so Claude has concrete patterns to match.

The skill should let me give it any rough text
and return a rewrite in this voice.

What happens. skill-creator reads the memos, extracts what it calls “rhetorical fingerprints” — architecture, verbal DNA, tone calibration, structural patterns — and writes a skill with instructions plus embedded examples. Save it. Name it something memorable like howard-marks-voice or firm-writing-style.

Verify. New chat. Give it raw notes:

Here are rough notes on how the market performed
this quarter: [bullets]. Turn this into a newsletter
using the writing-voice skill.

Claude produces a draft that reads like the source author — headers, paragraph rhythm, even the folksy sign-off. Not perfect, but genuinely close. This is the skill that people don’t believe works until they see their own newsletter come back in their own voice.

Skill 3: Inbox triage

Email is universal. Make Claude surface what actually matters.

Prompt:

Create a skill using the skill-creator skill that
triages my inbox.

When I run this, go through my recent email and
surface the ones that need my attention.

Focus on:
- Emails sent directly to me (not cc'd)
- Threads I haven't responded to
- Messages from clients, my boss, or key stakeholders
- Anything with time-sensitive language

Filter out:
- Newsletters, especially AI newsletters
- Promotional email
- Long threads that don't concern me
- Banter threads with friends

Use a red/yellow/green format:
- Red: urgent, respond today
- Yellow: respond this week
- Green: FYI or worth a quick read

Include the sender, subject, a one-line summary,
and your reasoning for the color.

Context matters here. If you have specific clients or colleagues who should always register as high-priority, name them in the prompt. skill-creator will fold them into the skill.

Verify. In a new chat with Gmail or Outlook connectors enabled:

Triage my inbox.

You should get a ranked list with color tags. The first few runs will miss nuance — maybe it flags an automated calendar invite as yellow when it should be noise, or misses a client because their domain is unfamiliar. Edit with natural language:

Edit the inbox-triage skill. Calendar notifications
should always be filtered out as noise. Any email
from @[your-client-domain].com should default to
at least yellow.

Replace. Re-run. This iterate-and-replace loop is where the skill gets tuned to you.

Bonus: chain two skills together

Once you have a transcript parser skill and a proposal generator skill, you can chain them:

Process this Zoom transcript using the transcript
parser skill. Then use the proposal generator skill
to draft a follow-up email based on the objections
and pain points surfaced.

Claude invokes both, in order, and hands you a finished first draft of a sales email. This is where the ROI compounds — a workflow that used to take 30 minutes now takes one.

Common mistakes

• Skipping “use the skill-creator skill” in the prompt. Claude reasons its way to a skill with general intelligence and you get a weaker output. Say the words every time early on.
• Not checking the sidebar. If you don’t see “reading the skill-creator skill”, it isn’t firing. Fix the prompt.
• Editing without replacing. The new version is a preview until you click replace. People build a great v2, wonder why outputs haven’t changed, and realize hours later they never replaced.
• Skipping the sample input. For extraction skills (research analyzer, transcript parser), attach a real example in the creation prompt. The examples are what teach Claude the shape of the output.
• Building one giant skill. “A skill that handles all my analyst work” will not work. Each skill should have one job. Chain them later.
• Not sharing. Skills are portable text files. If you build a working skill for a teammate’s workflow, send them the .md file. That’s the whole point.

A mental model for what to build

Anthropic’s framing: things you consume vs. things you create. Both are skill territory.

• Consume: research reports, legal documents, meeting transcripts, emails, credit agreements, SIMs. Extract, summarize, classify.
• Create: proposals, NDAs, LP letters, side letters, client briefing books, contract drafts. Produce with consistent format and voice.

Start with one of each. The research analyzer and style guide above are a good first pair. Inbox triage is a universal third. Once those are working you’ll start seeing candidates everywhere.

Next up

Skills run when you ask. Scheduled tasks make them run without you — at 11 PM, at 7 AM before you start work, weekly on Fridays. That’s where Cowork stops being reactive and starts being a proactive teammate.