78% of developers in coding subreddits now prefer Claude for programming tasks. That stat, from a recent Snyk analysis, is less surprising than the reason: it is not just the model — it is the skill ecosystem that has grown around it.
Claude Skills are directories with a SKILL.md file that give the model structured, repeatable workflows. Think of them as SOPs your AI actually follows. They load progressively — name and description first (~100 tokens), full instructions only when needed — and they work across Claude Code, Cursor, Codex, and Gemini CLI.
Here are the eight worth knowing about right now.
1. Planning with Files — 13,410 ⭐
Repo: OthmanAdi/planning-with-files
The most-starred skill in the ecosystem, and for good reason. It solves the single biggest failure mode of agentic AI: context loss mid-task.
The skill creates three persistent markdown files — task_plan.md, findings.md, progress.md — and enforces rules: re-read the plan before major decisions, save findings every two operations, track failures to avoid repeating them. It also implements a 3-strike error protocol and session recovery.
Activates automatically for tasks requiring 5+ tool calls. If you are doing anything multi-step — refactors, API builds, migrations — install this first.
2. Web Quality Skills — 496 ⭐
Repo: addyosmani/web-quality-skills — by Addy Osmani
Six integrated skills from the Google Chrome team lead covering everything a production web app needs to pass:
| Skill | What it covers |
|---|---|
performance |
Loading speed, <1.5MB page budget, JS <300KB compressed |
core-web-vitals |
LCP, INP, CLS — with framework-specific fixes for Next.js, React, Vue |
accessibility |
WCAG 2.1, screen reader support |
seo |
Search engine optimization |
web-quality-audit |
Comprehensive across all categories at once |
3. HashiCorp Agent Skills — 303 ⭐
Repo: hashicorp/agent-skills (official, MPL-2.0)
Official Terraform skills covering HCL code generation, .tftest.hcl testing patterns (including mock providers in Terraform 1.7+ and parallel execution in 1.9+), module refactoring, and HCP Stacks. The style-guide skill alone will stop Claude from generating Terraform that passes review on the first try — which is the only thing that matters.
4. Full Delivery Workflow — 82 ⭐
Repo: levnikolaevich/claude-code-skills
50+ skills encoding the entire software delivery lifecycle with a clean phase numbering system: 001–002 research, 100–150 docs, 200–230 planning, 300–310 task management, 400–404 execution, 500–513 quality gates, 600–624 auditing.
The separation of concerns is the feature: the task executor (ln-401) implements without committing. The reviewer (ln-402) reviews without implementing. The test executor (ln-404) writes tests without touching production code. This enforces actual process discipline instead of Claude doing everything in one shot and hoping for the best.
5. Core Engineering Framework
Repo: dralgorhythm/claude-agentic-framework
Nine concise, methodology-focused skills: TDD with Red-Green-Refactor cycles, systematic debugging, safe refactoring patterns, dependency management, and data-to-UI rendering. Each is under 2,500 characters and language-agnostic. Install this when you want Claude to apply discipline, not just generate code.
6. GitHub PR Review Skill
Repo: aidankinzett/claude-git-pr-skill
Prevents Claude from posting review comments on public PRs without your approval. It drafts everything first — file path, line number, comment text, suggestions — shows you exactly what will post, then requires explicit confirmation before touching the PR. The skill explicitly calls out rationalizations Claude might use to skip this step ("User said ASAP," "Only one comment"). Worth installing before you let Claude anywhere near a public repo.
7 & 8. Snyk Fix + Snyk Learning Path
Repos: snyk/studio-recipes, snyk-labs/snyk-learning-path-skill
Snyk Fix runs scan → analyze → fix → validate in one workflow. You can target by type ("fix dependency vulnerabilities"), by specific Snyk ID, or by CVE. It fixes bottom-to-top to avoid line-shift bugs, reruns the scan after fixing, and rolls back if it introduces new issues. Requires Snyk MCP server configured in your Claude Code environment.
Snyk Learning Path detects your stack from package manifests, Dockerfiles, and IaC files, then builds a 6–10 lesson learning path from Snyk Learn's 130+ free lessons using your actual codebase as the exercise material.
The Security Warning You Actually Need to Read
Snyk's ToxicSkills research found prompt injection in 36% of skills tested and 1,467 malicious payloads across the ecosystem. Before installing anything: read the full SKILL.md and every script in the scripts/ directory. Check the source. Review the allowed-tools frontmatter.
Where to Start
Install planning-with-files today. Add one or two domain-specific skills matching your daily stack. The productivity difference on multi-step tasks shows up in the first session — not because the model got smarter, but because it finally has a memory and a process.
Source: Snyk — Top 8 Claude Skills for Developers [delivered by ai.paul/blog_post]