learn-claude-with-phoebe / Session 3 of 6
Learn Claude with Phoebe · Session 3 of 6

From SOP to skill

Every team has procedures written down somewhere - report formats, QA checklists, onboarding steps. Today you turn one of yours into a skill: a reusable capability Claude applies automatically, so the standard is followed every time without re-explaining it once.

🟡 Medium All teams Team leads especially 45 min live + self-study
0-3 · Welcome 3-18 · Concepts 18-40 · Build your skill 40-45 · Q&A
Part 0

Why this session exists

Sessions 1-2 made you good at asking. But you're still re-explaining the same standards every time: how we format the weekly report, what our QA checklist covers, how a client email should sound. A skill packages that knowledge once - and Claude pulls it in automatically whenever the task matches. This is the session where your team's expertise becomes infrastructure.

Live - presented in session Self-study - read after class ★ Build-along demo 1 official course covered in full
★ What you walk out with today One real SOP from your own work converted into a working, tested skill - plus a curated shelf of skills worth installing and the judgment to know when a skill beats a prompt, a project, or a subagent.
Part 1 · covers "introduction to agent skills"

What a skill actually is 8 min live

A skill is a folder with instructions inside. That's genuinely it - no code required. The cleverness is in when Claude reads it.

LiveAnatomy: a folder, a SKILL.md, and a trigger3 min
📁 weekly-report/ SKILL.md name: weekly-report description: when the user asks for the weekly report... + the how-to instructions references/ scripts/ (optional) The description is the doorbell: Claude scans it to decide when the skill applies The body loads only when triggered: your steps, format, rules - context-efficient
🔍 Click to zoom - skill anatomy: the description triggers, the body instructs

Two parts matter:

  • The description (in the frontmatter) tells Claude WHEN this skill applies - it's read all the time, so it must name the trigger words and situations precisely. Weak description = skill never fires. This is the #1 troubleshooting lesson.
  • The body holds the HOW - your steps, format, tone, rules, examples. It loads only when needed, which keeps Claude fast and context-efficient even with dozens of skills installed.
  • Optional extras: a references/ folder for templates and examples, scripts/ for code Claude can run. Multi-file skills split detail out so the main file stays lean.
LiveSkill vs prompt vs project vs subagent - which tool when3 min

Session 1 gave you prompts, Session 2 gave you projects. Here's the full decision map - this covers the official course's "skills vs other features" lesson:

You have...UseBecause
A one-off taskA 4D promptNo reuse, no setup - just describe well
A domain you work in daily (team, client, product)A projectStanding knowledge + instructions for a WHOLE area
A repeatable PROCEDURE with rules and a formatA skillAuto-applies whenever the task matches, works across chats and projects
A big task needing isolated, parallel work (engineers)A subagentSeparate context window, returns a summary - Session 6
Real world

The tell that you need a skill: you catch yourself pasting the same instructions for the third time, or a teammate's output doesn't match yours because they didn't know rule #4. The weekly report, the campaign brief format, the data QA checklist, the client-email tone rules - each is a skill waiting to be written.

LiveWhat makes a GOOD skill candidate2 min
  • Repeatable: happens weekly or more, or across many people.
  • Rule-based: you can write down what "done right" means - format, sections, tone, checks.
  • Currently inconsistent: different people produce different quality - the skill IS the standard.
  • Bad candidates: one-offs, pure judgment calls ("should we enter this market?"), anything whose rules change weekly.
The 30-second test If you could hand the procedure to a smart new joiner as one page and they'd get it 90% right, it's a skill. If they'd need a month of shadowing you, it's still a job.
Part 2 · the conversion method

SOP in, skill out 5 min live

The method is a pipeline: pick, draft, test, tighten, share. You'll run the whole thing in Demo 1.

1 Pick 2 Draft 3 Test 4 Tighten 5 Share repeatable + rule-based SOP Claude drafts the SKILL.md from your SOP run 3 real tasks: does it fire? follow rules? fix the description, add the missed rule team folder or plugin - one standard for everyone
🔍 Click to zoom - the SOP-to-skill pipeline
LiveDraft it WITH Claude - the converter prompt2 min

You don't write the SKILL.md by hand - you paste your SOP into this converter and review what comes back. (There's also an official skill-creator skill that interviews you interactively; the converter below works everywhere, today.)

★ The SOP-to-skill converterConvert this SOP into a Claude skill as a single SKILL.md file. Requirements: - YAML frontmatter with: name (short-kebab-case) and description. The description must state exactly WHEN to use the skill - name the trigger phrases, task types and file types, because it decides whether the skill fires. - Body: numbered steps in working order, the output format (with a filled example), quality rules as a checklist, and what to do in edge cases. - Keep the body under 150 lines. If detail overflows (long templates, many examples), note what should move to a references/ folder. - Flag anything in my SOP that is ambiguous or contradictory instead of guessing. My SOP: [paste your procedure - bullet points are fine]
Real world

A marketing team's "campaign wrap-up report" SOP was 14 bullet points in Confluence that nobody followed exactly. Twenty minutes of converter + testing later: a skill that produces the same 6-section wrap-up from any campaign's raw numbers. The intern's report now matches the director's - that's the point.

Self-studyTesting and troubleshooting - why skills don't fire2 min read

The official course dedicates a lesson to this, because everyone hits it. The failure ladder, in order of likelihood:

  • Vague description. "Helps with reports" fires never; "when the user asks for the weekly performance report, a campaign wrap-up, or says 'run the report skill'" fires reliably. Name the trigger words.
  • Trigger overlap. Two skills with similar descriptions compete - make each description name what it does NOT cover.
  • Bloated body. A 600-line SKILL.md gets skimmed. Keep the body lean; push templates into references/.
  • Wrong tool. If it should apply to EVERYTHING in a domain, you wanted project instructions, not a skill.
The test protocol Run three REAL tasks: one that obviously should trigger it, one phrased sideways ("can you do the usual Friday thing?"), one that should NOT trigger it. Fix the description until all three behave.
Self-studyMulti-file skills and sharing them with the team2 min read
  • Multi-file structure: SKILL.md stays the lean brain; references/report-template.md holds the full template; references/examples.md holds 2-3 gold-standard outputs; scripts/ holds anything executable (engineers). Claude loads extras only when the task needs them.
  • Sharing: on Claude.ai/Teams, skills can be shared to the workspace so the whole team gets the same standard; in Claude Code, commit the skill folder to the repo (.claude/skills/) and everyone who clones has it. Plugins bundle several skills for distribution.
  • Versioning: treat skills like documents with owners - when the SOP changes, the skill changes, one place, everyone updated. That's the governance win over tribal knowledge.
  • Validating third-party skills: read the SKILL.md before installing (it's just text), check what scripts it runs, prefer official or well-starred sources - the same Diligence habit as ever.
Part 3 · don't build what you can install

The skill shelf: install these first 5 min live

Before building, shop. The ecosystem already has polished skills for the boring-but-critical stuff.

LiveThe starter shelf - official and battle-tested3 min
Skill / packWhat it gives youWho needs it
docx · pptx · xlsx · pdf (official, anthropics/skills)Proper Word, PowerPoint, Excel and PDF creation and editing - real files, correct formattingEveryone
skill-creator (official)Interviews you and builds new skills - the meta-skillEveryone building skills
theme-factory (official)Consistent visual themes across generated docs and pagesMarketing, anyone shipping visuals
marketing-skills pack (community, ~34k★)Copywriting, ads, SEO, emails, launch playbooks - dozens of focused skillsMarketing & branding (Session 4 uses these)
awesome-claude-skills lists (community)Curated indexes of hundreds of community skills - shop here before buildingEveryone, quarterly browse
Install rule of thumb Official pack first, then community skills with real stars and readable SKILL.md files. Read before you install - a skill is instructions, and you can audit instructions in two minutes.
LiveKOL skills - borrowing an expert's brain2 min

The most underrated genre: skills that encode a known expert's METHOD, not a company procedure. The pattern is called a KOL (key opinion leader) skill - you write down how a master works, and Claude applies that discipline to your material.

★ Example: a Karpathy-style learning skill--- name: karpathy-build-to-understand description: Use when the user wants to DEEPLY understand a technical concept (an algorithm, a model, a system) rather than just use it - trigger on "understand deeply", "from scratch", "build to understand", or "karpathy mode". --- Apply Andrej Karpathy's build-to-understand method: 1. Strip the concept to its minimal form - what is the smallest version that still works? 2. Build it from scratch in plain Python (no frameworks), under 100 lines, every line explained. 3. Introduce one real-world complication at a time; show how the minimal version grows. 4. At each step, ask the user to predict what breaks before revealing it. 5. End with: "what the production version adds, and why you now understand it." Rules: no framework imports until step 5. No hand-waving - if a step can't be coded simply, decompose it further. Spelled-out math over notation.

The same pattern works for any master you study: an Ogilvy-style ad-critique skill, a Tufte-style chart-review skill, a Buffett-style annual-letter summarizer. Write the method once, apply it forever - and yes, this is exactly how you'd package YOUR OWN expertise for the team as you become the in-house KOL.

Demo 1 of 2

Turn YOUR SOP into a working skill ★ 15 min · everyone builds

Bring a real procedure - the weekly report, the QA checklist, the client-email rules. By the end it's a tested skill.

Write your SOP as 5-10 rough bullets if it isn't written down already (3 quiet minutes - "how I actually do the Friday report").

Run the converter prompt (Part 2) with your SOP pasted in. Read the draft SKILL.md it returns - check the description names YOUR trigger phrases.

Answer the flags: the converter lists ambiguities it found in your SOP - resolving them now is free; resolving them after three bad reports is not.

Install it: Claude.ai → Settings → Capabilities → Skills → create/upload your skill (in Claude Code: save the folder under .claude/skills/).

Run the test protocol: one obvious task, one sideways-phrased task, one should-not-trigger task. Tighten the description until all three behave.

Victory lap: run your real task for this week with the skill live. Compare it with last week's manual version.

Data tip SOPs often contain client names and internal figures. Genericize the examples inside the skill ("Client A", rounded numbers) - the skill teaches the FORM, your live task supplies the real data.
Demo 2 of 2

Install from the shelf and run it ★ 7 min · everyone installs

Prove to yourself the ecosystem works: install one official skill and use it in anger.

Pick your poison: pptx if you make decks, xlsx if you live in spreadsheets, docx for documents.

Install it from the skills directory (Claude.ai → Settings → Capabilities → Skills → browse; engineers: anthropics/skills on GitHub).

Run a real job: "Create a 5-slide deck summarizing [this week's real topic] - our template style: title, 3 content slides, next steps." Watch it produce an actual .pptx.

Notice what you did NOT do: explain PowerPoint mechanics. The skill carried the how; you supplied the what. That division of labor is the whole session in one sentence.

After the session

This week ◐ 30 min total

Source material

Official courses covered

This session teaches the complete working content of the official skills course from claude.com/resources/courses.

Introduction to agent skills30 min · all 6 lessons: what skills are, first skill, config & multi-file, skills vs other features, sharing, troubleshooting
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Your skill never fires when you ask for the weekly report. First suspect?

The description is the doorbell. "Helps with reports" fires never; naming the exact phrases and task types fires reliably.

2 · You want Claude to know your team, product, and glossary in EVERY conversation. That's a job for...

Domain-wide context = project. A skill is for a repeatable PROCEDURE with rules and a format. Different tools, different jobs.

3 · Which of these is the best skill candidate?

Repeatable + rule-based + currently inconsistent - the skill IS the standard. Judgment calls and one-offs stay human work.

Session 3 cheat sheet · pin this

A skill isA folder with SKILL.md: description = when it fires, body = how it works. No code needed.
Which toolOne-off → prompt · domain → project · procedure → skill · isolated big task → subagent.
Good candidateRepeatable + rule-based + currently inconsistent. One page = skill; needs shadowing = still a job.
Not firing?It's the description. Name the exact trigger phrases and what it does NOT cover.
Test protocolObvious task · sideways phrasing · should-NOT-trigger. All three must behave.
Shop firstdocx/pptx/xlsx/pdf + skill-creator official; marketing pack for Session 4. Read before installing.