Skip to content
Agent Skills vs Prompt Engineering: What 164K GitHub Stars Are Telling You
TutorialsJuly 26, 202610 min read

Agent Skills vs Prompt Engineering: What 164K GitHub Stars Are Telling You

Agent skills (SKILL.md) replace system prompts with on-demand instruction files. Why 164K GitHub stars signal the shift and how to start your skill library today.

Agent skills are modular SKILL.md files that load on demand instead of bloating your system prompt. A typical skill costs 50-100 tokens in context, encodes one workflow class precisely, and works across Claude Code, Codex, Gemini CLI, and six other tools. Five of the top eight GitHub trending repos today are skills repositories.

I've been running a skills-based agent setup since November 2025. The consistency gap between a skill-driven workflow and a monolithic-prompt workflow is wide enough that going back isn't something I consider anymore.

What Is an Agent Skill?

An agent skill is a SKILL.md file inside a .claude/skills/ directory that teaches your agent how to handle a specific task class. The agent reads skill descriptions at startup (50-100 tokens each), then loads the full skill only when you invoke a matching task -- keeping context lean while giving the agent precise procedural knowledge exactly when it needs it.

Anthropic introduced the SKILL.md format in October 2025 and released it as an open standard in December 2025. By mid-2026 it was supported by eight major AI coding tools: Claude Code, Claude.ai, the Claude API, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot, Antigravity, and Windsurf. That cross-platform adoption is unusual -- most AI tool formats don't jump to competing products within a year of launch.

Matt Pocock's public skills library is the best reference implementation available right now. He published his personal .claude/skills/ directory -- the skills he actually uses for TypeScript tooling work -- and it crossed 164,000 GitHub stars in July 2026, gaining 1,740 stars in a single day on July 26. The star velocity tells you something: developers aren't just reading it for inspiration. They're copying it wholesale into their own setups.

Community

Stop building alone.

Join the Skool community. Ask questions, share what you're building, and learn from other people actually shipping AI agents.

Join Skool →

Why Prompt Engineering Breaks at Scale

Prompt engineering fails predictably when a single agent handles more than three or four task classes. The system prompt grows with each workflow until it consumes your context budget, becomes unmaintainable, and produces inconsistent output -- the model is weighting 5,000+ words of competing instructions against every single task simultaneously.

The specific failure modes show up fast. A monolithic prompt written for code review leaks review instructions into your documentation tasks. Instructions tuned for one developer's style don't compose cleanly when a second developer joins. Updating one workflow requires auditing the entire prompt to avoid unintended conflicts with unrelated instructions. There's no way to share just your "TDD workflow" with a colleague without handing over your entire prompt infrastructure.

Skills solve this by separating stored workflow knowledge from active task context. With 15 installed skills, you add roughly 750-1,500 tokens at startup -- the description index that helps the agent match tasks to skills. Then you load one full skill when needed, typically 500-2,000 tokens. Compare that to a 6,000-word system prompt front-loaded on every single task, regardless of whether those instructions apply.

The scalability difference is material. A skills-based agent handles 25 distinct workflow types without any one task being degraded by instructions meant for other tasks. A prompt-based agent covering 25 workflows either runs a bloated monolith or requires a human to manually paste the right context every time.

How the SKILL.md Format Works

A SKILL.md file has two parts: YAML frontmatter that declares the skill's name and trigger conditions, and a Markdown body with the step-by-step instructions the agent follows when the skill is active. The agent reads frontmatter at startup to build its skill index, then loads the full body when user input matches a trigger phrase.

Here's what a clean, working skill looks like:

---
name: code-review
description: Code review. Triggers on "review this", "check my PR",
"review the diff". Reviews for correctness bugs then simplification.
---

# Code Review

Run this process before reporting any review complete.

1. Read the full diff before commenting on any single line
2. Check for correctness bugs: null handling, off-by-one, race conditions
3. Check for simplification: shorter path to the same result?
4. Report findings grouped by: blocker / suggestion / nit
5. Never flag stylistic changes without a functional reason

When you say "review my PR", the agent loads this skill and follows the process. The review is consistent across sessions, across team members who share the same skill file, and across Claude Code updates that periodically change default behavior. The skill encodes your team's standards, not the model's defaults of the week.

Matt Pocock's library shows what a mature skill collection looks like in practice. His includes TDD as a skill (test first, fail first, minimal implementation), vertical slicing (break features into thin vertical slices rather than horizontal layers), TypeScript strict-mode enforcement, and triage -- deciding what to tackle next when a bug report is vague. Each one sharp, single-purpose, and independently testable. His working rule: if you can't describe the skill in one sentence, split it into two.

Want the templates from this tutorial?

I share every workflow, prompt, and template inside the free AI Creator Hub on Skool. 500+ builders sharing what actually works.

Join Free on Skool

What Makes a Good Skill (vs a Bad Prompt)

A good skill is single-purpose, has explicit trigger conditions, and encodes process rather than personality. A bad skill is a repurposed system prompt -- multi-scope, vague triggers, instructions that bleed into other workflows. If you can't describe what the skill does in one sentence, split it.

The patterns that consistently produce reliable skills:

Named triggers. List exact phrases that invoke the skill. "Triggers on: 'code review', 'check my PR', 'review the diff'" is precise. "Used when reviewing code" is ambiguous -- the agent fires it on tangentially related tasks. Trigger specificity correlates directly with skill reliability.

Process steps, not personality traits. "Read the full diff before commenting on any line" is a process step. "Be thorough" is a personality trait that produces different outputs every session. Skills that encode process are consistent. Skills that encode personality are not.

Explicit output format. Every skill should specify what the agent produces. "Report findings grouped by: blocker / suggestion / nit" is a target shape. Without it, output format varies by model version, context length, and whatever the agent decides is most helpful this particular session.

No cross-skill dependencies. Each skill should work standalone. If skill A requires skill B to be loaded, you've created implicit coupling that breaks when someone installs just one of them, or when you update skill B without realizing skill A depends on its exact behavior.

The VoltAgent/awesome-agent-skills repository on GitHub has 27,000+ stars and 1,000+ skills from real engineering teams. Its README explicitly excludes "mass AI-generated stuff" in favor of skills that teams actually use in production. Browsing it for 30 minutes teaches more about skill granularity than most written guides, including this one. The patterns in that collection are how experienced teams actually structure their skill libraries.

The Skills Ecosystem in Numbers

The adoption data is concrete. mattpocock/skills crossed 164,000 GitHub stars in July 2026, gaining 1,740 in a single day. The ECC superpowers collection has 233,000+ total stars. ComposioHQ/awesome-claude-skills added 577 stars in a single day this week. VoltAgent/awesome-agent-skills holds 27,000+ stars with 1,000+ production-tested skills from real engineering teams.

The broader signal: five of the top eight GitHub trending repositories on July 26, 2026 are skills-related. The community has been voting with stars and forks for months. The Towards AI piece on Claude Code's creator frames it directly: prompt engineering is no longer the main skill. Designing and maintaining a library of agent skills that encode your team's workflow standards persistently and composably -- that's the main skill now.

Cross-platform adoption confirms this isn't a single-vendor story. The SKILL.md format is now supported natively by Claude Code, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot, Antigravity, Windsurf, and OpenCode. A skill you write today for Claude Code works in Codex tomorrow with minimal or no modification. That portability has never existed for prompt engineering practices -- they've always been tool-specific and person-specific.

How to Start Building Your Skill Library Today

Start with one workflow you run at least twice a week. Not a complex one -- your simplest, most repeatable task. Write the SKILL.md to document exactly what you do, in order, with the specific checks you run and the output format you expect. Get one skill working reliably before building the second.

The practical starting sequence:

  1. Create the directory. Make .claude/skills/[skill-name]/SKILL.md in your project root for project-specific skills, or in ~/.claude/skills/ for global skills available across all repos. The agent discovers skills automatically -- no registration step required.
  2. Write the frontmatter. Name, one-line description, explicit trigger phrases. Keep the description under 100 words -- it loads at every session startup for every skill in your library, so it's part of your base context cost.
  3. Write the process. Number the steps. Specify what to check, what to produce, and what not to do. Use commands ("Read the full diff before commenting") not suggestions ("It's good to read the full diff"). Imperative is clearer.
  4. Test the trigger. Start a fresh Claude Code session, describe your task using one of the trigger phrases, and confirm the agent loads the skill and follows the steps in order. If it doesn't trigger, add more specific phrases to the frontmatter description.
  5. Iterate after real use. After three or four actual uses, you'll find steps that were ambiguous, output formats that don't match your actual needs, and trigger phrases that fire on the wrong tasks. Fix them. The skill compounds across every future use, so the iteration investment pays back quickly.
  6. Share it. Submit to ComposioHQ/awesome-claude-skills or VoltAgent/awesome-agent-skills. A skill that improves five teams' workflows compounds faster than one you keep private. The open-standard SKILL.md format was designed for exactly this kind of community sharing -- which is why it spread to eight tools in under a year.

If you want to see a mature library before building from scratch, clone mattpocock/skills on GitHub. Read three or four skills end-to-end. The pattern clicks quickly: these aren't complicated documents. They're precise ones.

FAQ

What is the difference between an agent skill and a system prompt?

A system prompt loads for every interaction and tries to encode all workflows simultaneously. An agent skill loads on demand when a task matches its trigger conditions. With 15 skills installed, you add 750-1,500 tokens of description-index overhead at startup, then load one skill (500-2,000 tokens) when needed -- versus front-loading 6,000+ words on every single task regardless of relevance.

Are agent skills the same as CLAUDE.md or AGENTS.md files?

No. CLAUDE.md and AGENTS.md load automatically for every session in a repository -- they are project-scoped instruction files covering context and conventions. Agent skills (SKILL.md) load on demand for specific task classes. They work together: CLAUDE.md sets project context, skills handle specific repeatable workflows without loading those instructions into unrelated sessions.

Can I use SKILL.md files with tools other than Claude Code?

Yes. The SKILL.md format is an open standard supported by Claude Code, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot, Antigravity, Windsurf, and OpenCode as of mid-2026. Skills written for Claude Code require minimal or no modification to run on other supported tools -- genuinely portable across the AI coding ecosystem in a way prompt engineering practices have never been.

How many skills should I have before it becomes a problem?

Practical experience suggests 15-30 skills per project is the sweet spot. At 50-100 tokens per description, 30 skills adds 1,500-3,000 tokens of startup overhead -- acceptable. Beyond 30, trigger disambiguation gets noisier and the agent occasionally loads the wrong skill for overlapping task descriptions. Better to have 20 sharp, well-scoped skills than 50 overlapping ones.

Want the templates from this tutorial?

I share every workflow, prompt, and template inside the free AI Creator Hub on Skool. 500+ builders sharing what actually works.

Join Free on Skool
AI Agents First

The daily signal from the frontier of AI agents.

Join builders, founders, and researchers getting the sharpest one-email read on what's actually shipping in AI — every morning.

No spam — unsubscribe anytime