Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

Prologue — A Skill Is Not Free
1 · The 2026 Claude Skills Marketplace Map
2 · Code Review and PR Workflow
3 · Release and Changelog
4 · Content Drafting
5 · Runbook Execution
6 · Research Helpers
7 · Codebase Exploration
8 · Personal Productivity
9 · Partner Integration Skills
10 · Curation Philosophy — Context Budget Accounting
11 · Role-based Skill Bundle Recommendations
12 · Marketplace Future — H2 2026 Outlook
Epilogue — Checklist, Anti-patterns, What's Next
References

Prologue — A Skill Is Not Free

The last post built a Skill from scratch — SKILL.md, metadata, example input/output, allowed-tools, folder structure. Its message: "if you need it, build it."

This one is the other side. Plenty of well-built Skills already exist. The Anthropic official marketplace, partner integrations, community registries — hundreds of Skills are out there. Should you install them all? Absolutely not.

A Skill is not free. Each installed Skill costs you:

Context window real estate (at minimum, SKILL.md metadata is always in the system prompt)
Decision tax — the agent has to evaluate "should I use this?" every turn
False-trigger risk — wrong Skill fires, wrong answer
Maintenance — when the original author moves on, the Skill goes stale

So the harder question is not "what to install" but "what to leave off." 200K of context is not infinite. Activate thirty Skills and the model can barely see your own code.

This post is curation. As of May 2026, which Skills earn their context cost? — by category. We close with role-based bundles for Backend, Frontend, SRE, PM, and Technical Writer.

1 · The 2026 Claude Skills Marketplace Map

First, where things live. Skills come from four broad sources.

1-1. The Anthropic Official Marketplace

The anthropic-skills plugin — official Skills built and maintained by Anthropic. In 2026 the core lineup is:

docx / pptx / xlsx / pdf — generate, edit, and read Office documents. The most battle-tested route to Word/PowerPoint/Excel/PDF.
skill-creator — a meta Skill for building or editing other Skills. Nearly required when writing your first one.
consolidate-memory — tidy up bloated memory files. Call it when CLAUDE.md balloons.
setup-cowork — helper for shared team context setup.

Rule of thumb for this bundle: "if in doubt, turn it on." High trust, often the foundation other Skills lean on.

1-2. Partner Integration Skills

Enterprise SaaS vendors built these in partnership with Anthropic. OAuth flow is built in; an MCP server typically comes along.

Atlassian (Jira, Confluence): create/search tickets, author wiki pages
Figma: read design files, extract components, generate design handoff specs
Canva: create and edit design assets
Stripe: read-only payment and subscription data — transactional execution is intentionally blocked
Notion: read/write pages and databases
Zapier: trigger Zaps, fanning out to thousands of integrations through one bridge
Slack: channel messages, DMs, thread work
Linear / Asana / ClickUp / Monday: project management integrations
HubSpot / Salesforce-style CRM: leads, deals, contacts
GitHub / GitLab: PRs, issues, code workflows

Match partner Skills one-to-one with the SaaS you actually use. A Skill for a tool you never touch is a false-trigger farm.

1-3. Community Marketplaces (GitHub-rooted)

Curation repos like awesome-claude-skills, individual maintainers' Skill repos, and growing "Skill collection" plugins. Notable 2026 sources:

The Skills section of anthropic-cookbook — official examples
awesome-claude-skills — community curation index
Maintainer dotfiles-style Skills — GitHub stars become a trust signal

Key caution: community Skills demand provenance review. A polished SKILL.md can mask a Skill that requests broad bash permissions. Before installing, read the allowed-tools block and shell scripts yourself.

1-4. In-house Skills

The highest-value category. Encodes your company's workflows, conventions, and domain knowledge. External Skills can't substitute (think "our PR template," "our incident response procedure"). Typically managed in an internal Git repo and symlinked into .claude/skills/.

This post focuses on external-Skill curation. Building in-house Skills was the previous post.

2 · Code Review and PR Workflow

The category most engineers turn on first. Effects are visible, so ROI is easy to read.

2-1. PR Summarizer

What: reads the PR diff and produces a human-readable "what does this PR do?" summary, usually injected as the first paragraph of the PR body.

Why it earns its cost:

Helps reviewers grasp the PR in the first thirty seconds.
Captures "why," which raw diff can't show.
For multilingual teams, automatic English summaries unlock global review.

Pairs well with: conventional commit drafter (below), security review Skill. Run all three on one PR and you get "summary + commit cleanup + security check."

When to install: teams whose PR bodies are often a single line (fix bug); teams shipping 5+ PRs per day.

When not to: solo projects; teams whose culture insists on hand-written PR bodies (redundant).

2-2. Security Review Skill

What: pattern-matches common security issues — SQL injection, XSS, secret leak, IDOR patterns, weak crypto — and drafts comments with severity tags.

Why: human reviewers can't keep security loaded every turn. "Looks fast" once, "add a test" once, then security slides. The Skill doesn't slide.

Pairs well with: PR summarizer, SCA tools like Aikido / Snyk via MCP. Skill is the first net; SCA is the second.

When to install: finance, healthcare, B2B SaaS — any domain where security is a business risk.

When not to: internal demos and prototypes. False-positive fatigue eats the value.

2-3. Conventional Commit Drafter

What: reads staged changes (git diff --staged) and proposes 2-3 Conventional Commits-compliant messages — type, scope, and a body that captures the "why."

Why: good commit messages are a resource for future debugging, release notes, changelogs, and git bisect. Writing "why" is what humans get laziest about — and what Skills help with most.

Pairs well with: release notes generator (below). Clean commit history lifts release-note quality automatically.

When to install: teams that adopted (or want to adopt) Conventional Commits.

When not to: solo projects where you already do this by reflex.

2-4. PR Checklist Enforcer

What: parses your internal PR template (Tests added? Migrations? Rollback plan? Monitoring?) and fills it in or flags blanks.

Pairs well with: in-house deploy-checklist Skill, in-house ADR Skill.

When to install: any team with a long PR template that is regularly left half-empty. One of the best ROI Skills on this list.

3 · Release and Changelog

3-1. release-notes-from-git-log

What: reads commits from the last tag to HEAD and produces user-facing release notes (### Added / ### Fixed / ### Breaking). Accuracy spikes when paired with Conventional Commits.

Pairs well with: conventional commit drafter (upstream), in-house changelog format Skill, social copy Skill (below) — release notes auto-translate into tweet and LinkedIn drafts.

When to install: teams shipping at least every two weeks. The hand-squeezing of release notes is one of the most statistically wasted blocks of engineering time.

When not to: "continuous deploy, no notes" cultures.

3-2. SemVer Decider

What: compares the change set against a public-API snapshot and recommends the next release as patch, minor, or major. Flags breaking-change candidates separately.

Why: SemVer is a perpetual library-maintainer headache. A wrong minor release breaks downstream builds.

Pairs well with: public-API extraction Skill, breaking-change detector.

When to install: library, SDK, public-API maintainers.

3-3. Changelog Migration Helper

What: Keep a Changelog format, GitHub Releases markdown, internal wiki — formats the same change information differently for each channel. Write release notes once, replicate three places.

Pairs well with: Notion integration, Slack channel auto-post Skill.

4 · Content Drafting

Engineers do not avoid writing because they cannot — they avoid the blank page. Skills fill the blank page for you.

4-1. Blog Post Drafter

What: takes input sources (code changes, release notes, meeting notes) and produces a blog post draft following your in-house tone and structure guide.

Pairs well with: brand voice Skill (below), Notion integration (auto-saves the draft).

When to install: developer marketing, DevRel teams, or any engineering org that runs a blog.

When not to: solo bloggers whose authenticity is the asset (over-reliance risk).

What: turns a chunk of release notes into a tweet thread, a LinkedIn post, an internal Slack announcement. Channel-aware tone, length, hashtags.

Pairs well with: release-notes-from-git-log (upstream). One pipeline: git log to release notes to social copy.

4-3. Brand Voice Enforcer

What: internalizes your tone and style guide (e.g., "use active voice," "trim exclamation points," "no empty marketing adjectives") and reshapes drafts to fit.

Why: with many contributors, voice drifts. People can't keep the style guide in their head. The Skill can.

Pairs well with: blog post drafter, marketing email drafter.

4-4. Meeting-to-Blog Converter

What: takes raw transcripts from tools like Granola or Fireflies and produces a publishable blog post draft. Sensitive-info masking + tone fix.

Pairs well with: Granola Skill, brand voice enforcer.

5 · Runbook Execution

SRE, platform, on-call territory. A single mistake is expensive, which is exactly why the Skill ROI is highest here.

5-1. Incident Triage Skill

What: ingests alert payloads (PagerDuty/Datadog/Sentry) and outputs (a) severity classification, (b) blast-radius estimate, (c) similar past incidents, (d) the first response steps.

Why: the first five minutes of an incident matter most and go wrong most. Showing a half-asleep on-caller "1) N users impacted, 2) similar to PD-1234, 3) first step: check A/B" is the real value.

Pairs well with: PagerDuty Skill, Sentry Skill, in-house runbook index Skill.

When to install: teams with 24/7 operational responsibility.

5-2. On-call Response Templates

What: "first user-facing response — collect info, stay calm," "first status-page update," "first internal-incident-channel message" — channel-appropriate templates filled in to the situation.

Pairs well with: Slack Skill, Statuspage MCP.

5-3. Runbook Execution Guide

What: indexes your internal runbooks (usually Markdown) and, when someone says "we need to do X," locates the runbook and walks through it step by step. Asks for confirmation before risky steps.

Pairs well with: AWS / GCP / Azure CLI MCP (below), Kubernetes Skill.

When to install: teams whose runbooks are actually maintained. If runbooks are stale, the Skill confidently lies.

5-4. Postmortem Drafter

What: receives raw incident timeline material (Slack thread, PagerDuty events, deploy logs) and drafts a blameless postmortem. Includes "5 Whys" and "Action Items" sections.

Pairs well with: Slack Skill, PagerDuty Skill, Notion Skill.

6 · Research Helpers

Where technical decisions are sourced. Misused, a time sink; well used, a time saver.

6-1. Paper Summarizer

What: takes an arXiv PDF or conference paper PDF and produces a five-section summary — "1) problem, 2) core idea, 3) results, 4) limitations, 5) one-line summary." Math is explained in prose.

Why: hundreds of AI papers a week. No one reads them all. We need a tool that, in 90 seconds, decides whether a paper is worth deeper reading.

Pairs well with: PDF Skill (official), Notion notes integration.

When to install: research engineers, applied ML teams, technical decision-makers.

6-2. Three-URL Comparison Table

What: takes three competitor product pages (e.g., Drizzle / Prisma / Kysely) and produces a feature comparison table. Marketing hype gets cut; rows are factual.

Pairs well with: web fetch tools, in-house ADR Skill.

When to install: teams that make frequent technology choices or evaluate vendors.

6-3. RFC Comparison Skill

What: takes 2-3 RFC or ADR documents and produces a "common / different / open questions" table.

Pairs well with: Notion / Confluence Skill.

7 · Codebase Exploration

When you join a new codebase or start a large refactor, the most expensive cost is "I don't know where anything lives."

7-1. Architecture Mapper

What: sweeps the codebase, builds a module dependency graph, traces data flow, lists external integration points, and lists the system's "five hot spots." Markdown out.

Why: the most-needed onboarding artifact in week one. Drawn by hand, always stale.

Pairs well with: Mermaid diagram generator, in-house ADR Skill.

When to install: large monoliths, dozens of microservices, or inheriting legacy code.

7-2. Dead Code Finder

What: uses import-and-call graph analysis to flag "functions, files, endpoints nothing else calls." Surfaces dynamic-import / reflection cases as possible false positives.

Why: dead code taxes every future reviewer's attention. The difference in codebase health between teams that clean regularly and teams that don't is stark by year two.

Pairs well with: test coverage Skill, in-house refactor Skill.

When to install: any codebase older than two years.

7-3. Dependency Auditor

What: reads package.json, pyproject.toml, go.mod and lists (a) unused deps, (b) security advisories, (c) major-version upgrade candidates.

Pairs well with: Aikido / Snyk MCP.

7-4. Test Gap Analyzer

What: finds code paths not exercised by tests and ranks them by "how business-important is this path?" to prioritize new tests.

Pairs well with: coverage tools (vitest --coverage, pytest-cov), in-house critical path definitions.

8 · Personal Productivity

The category that most directly buys back your time.

8-1. Daily Standup Writer

What: aggregates yesterday's PRs, commits, closed tickets, and tomorrow's calendar into a "yesterday / today / blockers" note. Auto-posts to Slack.

Pairs well with: GitHub Skill, Linear / Jira Skill, Slack Skill, Calendar Skill.

When to install: teams that run async daily standups.

8-2. Meeting-to-Action-Item Extractor

What: consumes a raw meeting transcript and extracts just the action items, by owner and due date. After your confirmation, files tickets to Linear or Jira.

Pairs well with: Granola Skill, Linear Skill.

8-3. Email Triage Helper

What: sweeps your inbox and categorizes — "answer today / this week / FYI / can ignore." For the simple ones, drafts a response.

Pairs well with: Gmail or Outlook integration.

When not to: roles that barely use email. Context window waste.

8-4. Weekly Retro Writer

What: aggregates the week's PRs, commits, meeting notes, and calendar into a three-section retro — "shipped / didn't / next week priorities."

Pairs well with: GitHub Skill, Calendar Skill, Notion Skill.

9 · Partner Integration Skills

SaaS integrations usually arrive as an MCP-server + Skill pair. MCP provides the tools; the Skill provides the workflow knowledge.

9-1. Atlassian (Jira / Confluence)

What: search, create, move tickets; read and write wiki pages. The Skill knows workflow patterns like "how to write a good bug ticket" and "how to author a release notes page."

When to install: any team using Jira or Confluence as its source of record.

9-2. Figma

What: extract components, styles, and assets from design files; auto-generate design-dev handoff specs; convert design tokens to code tokens.

Pairs well with: Storybook integration, design system Skill.

When to install: frontend or full-stack teams; design system operators.

9-3. Canva

What: generate design assets — social cards, presentation slides, one-page specs. Text-in design-out.

When to install: marketing teams, DevRel.

9-4. Stripe

What: read payment, subscription, invoice data. Transactional execution is blocked — an intentional safety. "Why did customer X's last payment fail?" type queries.

When to install: anyone who reads Stripe regularly — CS, Finance, engineering on-call.

9-5. Notion

What: read and write pages and databases. The Skill knows workflows like "auto-summarize meeting notes," "auto-index project docs."

Pairs well with: Granola, Calendar.

When to install: teams whose wiki is Notion.

9-6. Zapier

What: triggers Zapier's thousands of integrations through Zaps. "Create a Salesforce lead," "add Mailchimp subscriber" exposed as Skills.

Why it's valuable: instead of installing every lightweight SaaS Skill individually, one Zapier bridge covers them. Caveat: response latency is higher.

When to install: marketing or sales teams that need broad lightweight SaaS integrations.

9-7. Slack / Linear / GitHub

The productivity triad. Almost every engineering team's baseline. Each Skill is well-tested and stable.

10 · Curation Philosophy — Context Budget Accounting

Now the hard part. Every Skill you install spends part of your context. This is not abstract.

10-1. The Three Real Costs of a Skill

Metadata cost — SKILL.md's description sits in the system prompt every turn. A 200-character description × 30 Skills = 6,000 characters always loaded.
Decision tax — more Skills means more tokens spent on "which one do I use?" Latency rises.
False-trigger cost — the wrong Skill firing produces wrong answers. Users pay in debugging time.

10-2. ROI Matrix — Which Skills Justify Their Cost

Category	Use frequency	Task value	Justified?
Incident triage	Low (1-2x/week)	Very high (revenue at risk)	Always on
PR summarizer	Very high (daily)	Medium	On
Paper summarizer	Medium (5-10x/week)	High	On
Meeting converter	Medium	Medium	Conditional
Email triage	Role-dependent	Low-to-high	Role-dependent
Stripe Skill	Low	High (money)	CS/Finance only

The rule is simple: "will I use this even once this week?" — if "probably not," off. Anything you keep on should trigger at least once per week on average.

10-3. The On vs Off Workflow

Schedule a 30-minute Skill audit each quarter. Look at which Skills actually fired during the quarter.
Off any Skill with zero uses. "Just in case" is a bad reason.
If false-trigger frequency is high, tune the description or turn it off.
Prioritize in-house first, community second, partner third, generic official last. (Higher company specificity ranks higher.)

10-4. Five Anti-Patterns

"Install the entire marketplace." A 30-Skill environment is always worse than a 1-Skill environment.
"Chase trends." Don't just add a new Skill because the beta launched. Try for a week, then evaluate quantitatively.
"Trust community Skills without review." If you haven't read the allowed-tools block and any shell scripts, don't install.
"Ignore description collisions." When two Skills' descriptions are similar, false triggers are inevitable. Keep one, or clarify the descriptions.
"Use partner Skills as in-house substitutes." Your PR template is unknown to external partners. Company-specific workflows belong in in-house Skills.

11 · Role-based Skill Bundle Recommendations

Based on the above, suggested bundles by role. These are starting points, not final answers — adjust to your team.

11-1. Backend Engineer (under 15)

Core: code review + release + daily workflow.

Official: skill-creator, consolidate-memory
Code: PR summarizer, conventional commit drafter, security review
Release: release-notes-from-git-log, semver decider
Explore: architecture mapper, dead code finder
Integration: GitHub, Linear/Jira, Slack
Personal: daily standup writer

11-2. Frontend Engineer (under 15)

Core: design handoff + component review.

Official: skill-creator
Code: PR summarizer, conventional commit drafter
Design: Figma Skill, design system Skill (in-house), brand voice
Content: changelog-to-social-copy
Integration: GitHub, Linear, Slack, Figma
Personal: daily standup writer

11-3. SRE / Platform Engineer (under 12)

Core: incidents + runbooks + automation.

Official: skill-creator, docx (for postmortems)
Runbooks: incident triage, on-call response, runbook guide, postmortem drafter
Integration: PagerDuty, Sentry, Slack, GitHub, AWS/GCP CLI MCP
Code: PR summarizer

11-4. Product Manager (under 10)

Core: meetings + research + writing.

Official: pptx, docx
Research: paper summarizer, three-URL comparison
Content: blog drafter, brand voice
Meetings: meeting-to-action-item extractor
Integration: Linear/Jira, Notion, Slack, Figma (read-share)
Personal: weekly retro writer

11-5. Technical Writer / DevRel (under 10)

Core: content pipeline.

Official: docx, pptx, pdf
Content: blog drafter, changelog-to-social-copy, brand voice, meeting-to-blog converter
Research: paper summarizer, three-URL comparison
Integration: Notion, Canva, Slack, GitHub (release notes)

11-6. New Hire / Onboarding (5-8)

Core: learning + exploration. Start small.

Official: skill-creator (you will build one eventually), docx
Explore: architecture mapper, dependency auditor
Code: PR summarizer
Integration: GitHub, Slack
Personal: daily standup writer

12 · Marketplace Future — H2 2026 Outlook

Briefly, where this is going as of May 2026.

Skill composability becomes the primary value driver. A graph of small Skills calling each other beats one giant Skill.
Partner Skills move to SLA models. Free Skills cannot guarantee context-cost worth, so users turn them off. Paid partner Skills evolve to "false-trigger rate under 1%" guarantees.
In-house Skill hosting services. Internal plugin registries replace dotfiles-style internal Skill management.
Skill audit logs. Compliance starts mandating "which Skill fired when, how." SOC 2 implications.
Skill curation as a job. "AI Skill Curator" emerges in larger orgs — manages the internal Skill catalog, vets external Skills, runs the quarterly audit.

Epilogue — Checklist, Anti-patterns, What's Next

Curation is essentially "subtraction," not "addition." The decision to turn a Skill off is harder and more valuable than the decision to turn one on. Schedule that 30-minute quarterly audit.

Starting Checklist

Pick the section-11 bundle that matches your role and run it for one week
Track which Skills triggered (Claude Code has automatic logs)
After one week, turn off any Skill that did not trigger
Tune or turn off Skills with high false-trigger rates
Calendar a quarterly 30-minute Skill audit
Company-specific workflows belong in in-house Skills (external Skills can't substitute)

Anti-patterns Quick Recall

"Install the whole marketplace" — 30-Skill environment is worse than 1-Skill
"Chase trends" — no add without a week of quantitative evaluation
"Trust community Skills blindly" — read allowed-tools and shells yourself
"Ignore description collisions" — similar descriptions breed false triggers
"Substitute in-house Skills with partner Skills" — external doesn't know your company

What's Next

Building an in-house Skill catalog — Git repo structure and permission model
Eval systems for Skills — false-trigger regression alerts
MCP servers vs Skills — what's a tool, what's a workflow
The "AI Skill Curator" role — a day in the life of the catalog manager

"A Skill is not free. It costs part of a finite resource — the context window. Deciding what not to install is harder and more valuable than deciding what to install."

— Claude Skills Marketplace 2026, end.