Skills — Make the Machine Remember How You Think

Day 1f · Skills

Make the machine
remember how you think.

Reusable instructions, the Anthropic skill library, when to build vs. borrow, and the one rule that saves you from yourself.

20 min Katherine Cowork + CLI Live demo included

The concept

A skill is a saved recipe.

Official definition

Reusable instructions saved as a named command. You write it once. Claude reads it when it becomes relevant — not before. No code needed.

/brief-writer

/deck-qa

/data-clean

/comms-triage

Namebrief-writer

DescriptionGenerates client campaign briefs. Use when starting a new campaign.

You provideClient name, campaign goal, budget

I produceObjectives, audience, key messages, channels

Tone rulesCompany voice. No jargon. Action-oriented.

Works in both: Claude Cowork (desktop app, click-to-use) and Claude Code CLI (type the command). Same concept, different surface.

A confession

I built 12 skills
I never used.

Automating before I understood the task.

What I did

Built skills for workflows I stopped needing once I actually understood the tools. Half of it was compensating for weak prompts.

What the Anthropic docs actually say

Build evaluations first. Run the task manually. Document specific failures. Build only enough to fix those gaps. Don't speculate on what you'll need.

Automating a broken process makes a faster broken process.

The one rule

Do it manually
2–3 times first.

1×

Rough draft.
Still exploring.

→

2×

Cleaner. You see the pattern.

→

3×

Same shape every time. It's working.

→

⚡

Now build the skill.

If it's still changing — it's not ready.

Degrees of freedom

Prescriptive or open-ended?

More prescriptive More open-ended

🔵 Lock it down

Faster, consistent output
Easy to review and trust
Great for tasks with right answers
Brittle — edge cases break it
Needs updating as formats change

Best for: report formatting, brand QA, data cleaning

🟢 Leave room

Adapts to varying context
Handles edge cases gracefully
Better for judgment-heavy tasks
Less predictable output
Needs more review

Best for: comms triage, ideation, client responses

From the Anthropic skill docs: "Think of Claude as a robot exploring a path — narrow bridge vs. open field."

Case study · /brief-writer

Same prompt. Two different skills.

New client: Gen Z fashion brand. TikTok-first. No budget set yet.

🔵 Locked down

• Output exactly: Client Overview / Objective / Budget: £X,XXX / 3 Key Messages • Tone: Always formal. No contractions.

CLIENT OVERVIEW: The client operates in the fashion vertical, targeting Generation Z consumers through digital channels. BUDGET: £[UNKNOWN] ← format breaks OBJECTIVE: To engage the youth demographic leveraging social platform activations...

😕 Tone wrong. Budget field broken. 15 min of edits anyway.

🟢 Flexible

• Write a campaign brief. • Adapt format and tone to the client type.

Gen Z Fashion Brief 🧵 Goal: TikTok-native awareness before the spring drop. Audience: 18–24, already on the algorithm. Tone: Bold, no cringe. Budget: TBD — nail the concept first.

😊 Reads the room. Actually useful.

The prescriptive version was perfect for your last 5 clients. One new scenario — no budget, wrong tone register — and it breaks. That's brittleness.

The plugin marketplace

Anthropic already
built some of this.

Try first
📊
Marketing
Campaign reporting, briefs, channel analysis

Try first
📧
Gmail
Read, triage, and draft replies from Cowork

💼

Sales

Pipeline, outreach, CRM workflows

📁

Google Drive

Read, summarise, edit docs in Drive

💬

Slack

Summarise threads, draft replies, monitor channels

⚖️

Legal

Contract review, clause flagging

👥

HR

Job descriptions, policies, onboarding

🔨

Plugin Create

Build your own plugin — Anthropic guides you

Cowork → Customize → Browse plugins · Available on Pro, Max, Team, Enterprise plans

For a marketing consultancy

Four skills worth building.
Check the marketplace first.

Weekly reporting /weekly-report

Data Pull + Commentary

Paste in raw numbers. Get formatted commentary with trend flags and client-ready language. Prescriptive — same shape every time.

🔵 Check Marketing plugin first

Survey: 1–2 hrs/week. High ROI.

Brief writing /brief-writer

Campaign Brief Generator

Describe client and goal. Skill fills in structure: objectives, audience, messages, channels, KPIs. Consistent format every time.

🔵 Check Marketing plugin first

Survey: 30–60 min/week. Easy to standardise.

Deck QA /deck-qa

Brand Compliance Check

Paste slide text. Checks headline length, passive voice, CTAs present, sources cited, brand terminology. Flags issues with line refs.

✓ Prescriptive — clear rules, right answers

Comms triage /comms-triage

Inbox Triage + Drafts ⭐

Paste emails or connect Gmail plugin. Categorises by urgency, drafts responses for routine items, flags escalations. Semi-open ended.

🔵 Gmail plugin handles this natively

Survey: 30–60 min daily. Demo up next →

Sharing with your team

A skill is just a text file.
Share it like a Google Doc.

From the Anthropic docs: The description field is critical — Claude uses it to choose the right skill from 100+ options. Make it specific. "Generates client campaign briefs" beats "helps with documents."

🎤 Live Demo

Marketing plugin · /performance-report · demo/ folder attached to project

"Your Monday morning hour."

Client approved AI use. Data already cleaned by the team. Folder attached. One command.

📁 Project · demo/ folder attached

📁 demo/ meridian-q1-campaign.csv meridian-context.md /performance-report W/E 28 Feb 2026

📝 Output (~28 sec)

Meridian Active · W/E 28 Feb 2026

Executive Summary

Email Win-Back (7.1x) and Meta Stories (4.4x) are carrying the account. Display Prospecting is in negative ROI — pause immediately.

What Worked ✅

• Email Win-Back — 7.1x ROAS, £1,100 spend. Underinvested.

• TikTok — +26% MoM (1.09→1.37x). Spring creative working.

What Needs Improvement 🚩

• Display Prospecting — 0.39x ROAS. Pause by Wednesday.

• Google Non-Brand — budget ↑ this week, ROAS ↓ 1.45→1.21x. Bid review.

Switch to Claude. Cowork → project with demo/ attached → type /performance-report W/E 28 Feb 2026 → send

Going further

Skill → Script → Scheduled job

Only move up the stack when the value justifies the maintenance.

/

Skill

Manual invoke
/comms-triage

⚙️

Script

Triggered by event
e.g. file drop

🕐

Scheduled job

Runs automatically
Monday 8am Slack report

Example progression

/deck-qa run manually today → cron job that runs on every new Figma export next quarter

🌱 Environmental note

Reusable automation uses far less compute than 30 manual conversations re-explaining the same thing. Efficient skills are the responsible answer.

✨ Bonus · Survey-matched finding

The highest-leverage use of AI for consultants

Claude as your
thinking partner.

🔬

Pressure-test

"Red team my brief." "What are the three weakest assumptions?" "What would a skeptical client say?"

💡

Deepen ideation

"What am I missing?" "Give me three angles I haven't considered." "Push harder on this insight."

🪞

Challenge assumptions

"Disagree with me." "What would the opposing view be?" "Where is this logic circular?"

From your survey · and from AI research

"Gaining insights from different sources to push ideation to a deeper level." — Verification and generation are distinct skills. You don't have to generate the idea to add enormous value. AI pressure-tests. You judge what matters to your clients.

The cheat sheet · Day 1f

If you remember nothing else,
remember these.

The one rule · write this down

Do it manually 2–3 times before you build a skill.

You can't automate what you don't understand yet.

📝

Skills = saved recipes

Write once in plain English, call with /slash-command. No code. Works in Cowork and CLI.

🛒

Check the marketplace first

Marketing, Gmail, Google Drive plugins already exist. Cowork → Customize → Browse plugins.

🔵🟢

Flexible beats rigid for client work

Lock it down only when format never varies (reports, QA). Leave room when clients and tone differ.

🪞

Use it to think, not just produce

"Red team my brief." "What am I missing?" Pressure-testing is where the real value is.

This week →

Pick one task you do every week. Do it with Claude twice more. If the output looks the same shape — build the skill. Still changing? Not ready yet.