Top 7 Claude Skills That Reduce Token Usage in 2026

Every token your agent burns is money, and most of that money is wasted on the wrong things: bloated context, agents that wander in the wrong direction, and skills that load when they shouldn't. The agent skills ecosystem has exploded in 2026, with indexes tracking over a million SKILL.md files, so the question isn't whether skills exist. It's which ones actually cut your token bill instead of inflating it. Some skills, like Matt Pocock's caveman, exist specifically to compress output. Others, like GSD Core, fight context rot by running work in fresh windows. And some platforms exist to help you verify which skills are worth installing in the first place, which matters because research found roughly 26% of published skills carry at least one vulnerability. This list covers seven options, starting with the one we think solves the verification problem best, then moving through the skills and directories that genuinely help you spend fewer tokens per task.

#1 SkillzBench — Editors' choice

SkillzBench is AIQ Labs' free, independent directory of AI agent skills, and it takes a position almost nobody else does: before you install a skill to save tokens, you should know whether it actually works and whether it's safe. The platform continuously crawls and dedupes skills from GitHub, then pins every version to a commit SHA, validates it, checks its license, security-scans it, and judges its quality. The part that matters most for a token-reduction listicle is the benchmark: SkillzBench tests every skill in a sandbox with the same model, with and without the skill, and publishes the resulting Lift in percentage points. That means you can see whether a skill that claims to cut token usage actually does, before you spend anything installing it. Everything on the platform is free: the directory, the benchmarks, the alerts, and a free MCP server that connects any agent to the skills you choose. Per-agent loadouts let you toggle skills and tools per agent, so each agent only sees what you picked, which itself keeps context lean. Trust scores run 0 to 100 and are built from security, quality, license, maintenance, and community signals, and money never touches scores.

Best for: Anyone installing agent skills who wants proof a skill helps before it touches their context window

Pricing: Free

Key features

• Continuous crawling and deduplication of skills from GitHub
• Every version pinned to a commit SHA, with validation, license checks, and security scans
• Sandbox benchmarks that test each skill with and without the skill, publishing Lift in percentage points
• Free MCP server that connects any agent to the skills you choose
• Per-agent loadouts so each agent only sees the skills and tools you toggle on
• Trust score from 0 to 100 based on security, quality, license, maintenance, and community signals
• Upstream tracking with re-scans, re-benchmarks, and notifications when skills change
• Everything free: directory, MCP connector, benchmarks, and alerts

Pros

• Benchmarks show measured Lift, so token-saving claims are backed by tests, not marketing
• Free MCP connector and per-agent loadouts keep each agent's context limited to what you chose
• Security scanning matters: published research found roughly 26% of skills carry at least one vulnerability
• Trust scores stay separate from Lift, and advertising money never influences scores or listings

Cons

• It's a directory and benchmarking layer, so you still install skills into your agent of choice
• Relatively new platform, so depth of coverage varies by category

Visit SkillzBench

#2 GSD Core (Git. Ship. Done.)

GSD Core is a spec-driven skill system that pitches itself as a fix for context rot, the slow drop in quality as your context window fills during a long session. That same mechanism is why it shows up on a token-reduction list: it breaks work into atomic plans that run in fresh context windows, keeping the main session light while subagents do the heavy lifting. Instead of one giant conversation accumulating hundreds of thousands of tokens, work happens in smaller, isolated bursts. According to the project's documentation and community coverage, it targets more than a dozen agents, and installation runs through an installer that prompts for your runtime (npx @opengsd/gsd-core@latest). It's worth being honest about the tradeoff, though. One developer who tested it reported that it burned through tokens fast and takes over your whole workflow, which is a fair warning for anyone evaluating it purely as a cost saver. Fresh subagent contexts reduce the per-task footprint, but the orchestration itself has overhead.

Best for: Developers running long, multi-step agent sessions who want to keep the main context window small

Pricing: Free and open source

Key features

• Atomic plans that run in fresh context windows
• Subagents handle heavy lifting while the main session stays light
• Explicitly designed to counter context rot in long sessions
• Installer supports more than a dozen agent runtimes
• Install via npx @opengsd/gsd-core@latest

Pros

• Directly targets context rot, one of the biggest hidden token costs in long sessions
• Works across many agent runtimes, so you're not locked to one tool
• Open source and free

Cons

• One tester reported it burned through tokens fast, since orchestration adds its own overhead
• It takes over your entire workflow, which some users dislike

Visit GSD Core (Git. Ship. Done.)

#3 Matt Pocock's skills (caveman, grill-me, and friends)

Matt Pocock's skill collection includes one skill that's explicitly about token costs: caveman, which compresses the agent's output style. The broader collection takes a deliberate stance against all-in-one systems, on the theory that approaches which own your whole process take away your control and make bugs hard to fix. So his skills are small, composable, and easy to adapt. grill-me makes the agent interview you before it writes anything, which front-loads a short conversation instead of a long, expensive wrong build. There's also grill-with-docs, which does the same interview and builds a shared-language document that pays off across sessions, plus tdd for a red-green-refactor loop and improve-codebase-architecture for rescuing spaghetti code. Install the ones you want with the Vercel skills CLI (npx skills@latest add mattpocock/skills). On the skills.sh leaderboard tracked in August 2026, grill-me sat at 756.3K installs and had overtaken Anthropic's frontend-design as the most-installed non-bundled skill.

Best for: Developers who want small, composable skills and explicit output compression

Pricing: Free and open source

Key features

• caveman skill specifically designed to reduce token costs
• grill-me interviews you with 10 to 15 questions before building anything
• grill-with-docs builds a shared-language document that pays off across sessions
• Small, composable skills instead of one monolithic workflow
• tdd skill enforces a red-green-refactor loop
• Install selectively via npx skills@latest add mattpocock/skills

Pros

• caveman directly targets token costs rather than implying savings
• Composable design means you only load what you need
• grill-me kills wrong-direction builds early, which is where most token waste happens
• Very popular: hundreds of thousands of installs tracked on public leaderboards

Cons

• grill-me can feel tedious when you just want something done quickly
• The interview itself consumes tokens before you save any

Visit Matt Pocock's skills (caveman, grill-me, and friends)

#4 Superpowers (obra/superpowers)

Superpowers is the most-starred skill collection in the ecosystem, sitting around 226,000 GitHub stars as of community coverage, ahead of the Vue repo. It's a full development methodology that triggers automatically: when the agent sees you're building something, it stops, works out a spec with you, gets your sign-off, writes an implementation plan, then hands the plan to subagents that build and review each other's work. The token angle is indirect but real. By forcing a spec and plan before any code gets written, it prevents the expensive failure mode of an agent charging down the wrong path and rewriting everything. That said, honest coverage is required here: the developer who reviewed it noted that it can eat up a lot of tokens, because the methodology itself involves multiple subagents and review passes. It's available through the official Claude plugin marketplace, plus Codex, Cursor, Gemini CLI, and others.

Best for: Teams that want one system owning the whole loop from idea to reviewed code

Pricing: Free and open source

Key features

• Full spec-driven development methodology that triggers automatically
• Spec and plan approval before any implementation begins
• Subagents build and review each other's work
• Available via the official Claude plugin marketplace, Codex, Cursor, and Gemini CLI
• Includes skills like Test-Driven Development, Systematic Debugging, and Verification Before Completion

Pros

• Prevents wrong-direction builds, which is the most expensive form of token waste
• Very large community and extensive documentation
• Works across many agents, not just Claude Code

Cons

• The methodology itself can eat up a lot of tokens, per developer testing
• Opinionated: it takes over your entire process
• Mostly redundant if your agent already has spec-driven workflows built in

Visit Superpowers (obra/superpowers)

#5 The Karpathy guidelines (multica-ai/andrej-karpathy-skills)

This one is almost the opposite of Superpowers: a short file of behavioral rules, now at roughly 174,000 stars. Despite the name, Andrej Karpathy didn't write it. It's a community-built file derived from a January post where Karpathy listed the ways LLM coding agents tend to fail, and the repo later moved to the multica-ai org. The rules come down to four ideas: think before coding, keep it simple, make surgical changes, and stay focused on the goal. The token-reduction case is straightforward. Agents that touch too many files and skip planning generate far more output than a disciplined agent making surgical changes. Because it's a single CLAUDE.md (or AGENTS.md, or usable as a Skill), the install is one curl command, and it adds almost nothing to your context. It's the low-commitment starting point if Superpowers feels like too much process. One honest caveat: the developer who tested it wasn't sure it was worth it in his own testing, so treat it as a cheap experiment rather than a guaranteed win.

Best for: Developers who want a minimal, low-commitment behavioral guardrail

Pricing: Free and open source

Key features

• Four core rules: think before coding, keep it simple, make surgical changes, stay focused
• Derived from Karpathy's public list of how LLM coding agents fail
• Single-file install as CLAUDE.md, AGENTS.md, or a Skill
• Minimal context footprint, one short file
• Install via a single curl command

Pros

• Tiny footprint: one short file adds almost nothing to context
• Surgical changes mean less generated code and fewer rewrites
• Easiest possible install, one command

Cons

• Effectiveness is debatable: at least one tester was unsure it was worth it
• No tooling, benchmarks, or measurement behind it, just rules

Visit The Karpathy guidelines (multica-ai/andrej-karpathy-skills)

#6 skills.sh

skills.sh is a CLI-installable registry and leaderboard for agent skills, and it earns a spot here mostly through the discipline it enables. Its public leaderboard shows install counts, which is how the wider community tracks which skills are actually being adopted, including the token-focused ones like caveman and grill-me. The install route is the Vercel skills CLI (npx skills add), which lets you add individual skills from a repo rather than cloning entire collections, so you keep your installed set lean. That matters for tokens: every loaded skill costs a little context, so selective installation beats bulk installs. Coverage of the platform notes it's an open, repo-based ecosystem with a developer-leaning audience, minimal curation, and free access. It's a discovery and distribution layer rather than a benchmarking one, so it won't tell you whether a skill reduces token usage, only whether lots of people installed it.

Best for: Developers who want to discover and selectively install popular skills from the command line

Pricing: Free

Key features

• Public leaderboard with live install counts across the skill ecosystem
• CLI installation via npx skills add
• Per-skill selective installs instead of whole-repository clones
• Cross-platform Agent Skills support
• Free to use

Pros

• Install counts give a rough signal of what the community actually uses
• Selective per-skill installs keep your context lean
• Fast, terminal-first workflow with no account friction

Cons

• Minimal curation, so quality and safety are on you to verify
• Install counts measure adoption, not whether a skill reduces token usage

Visit skills.sh

#7 Skills Directory (skillsdirectory.com)

Skills Directory is a large indexed collection of agent skills with a security-first stance. According to their website, they've indexed hundreds of thousands of skills and run automated security analysis on every listed one, scanning for prompt injection, credential theft, and data exfiltration across roughly 120 detection patterns and 10 threat categories. They also report that 36% of skills in the wild have security flaws, which is a sobering number for anyone bulk-installing skills to trim token costs. The token angle here is about avoiding waste and risk: a skill that exfiltrates data or injects instructions can cause an agent to burn tokens on unintended work, and a bloated or broken skill pollutes context for nothing. Installation runs through their CLI (npx -y skills find), and submissions are reviewed before publishing. Like skills.sh, it's a discovery layer, so it won't measure whether a skill reduces token usage for you.

Best for: Developers who want security-scanned skills before installing anything into their agent

Pricing: Free

Key features

• Automated security analysis on every listed skill
• Scans for prompt injection, credential theft, and data exfiltration
• Roughly 120 detection patterns across 10 threat categories
• CLI install via npx -y skills find
• Reviewed submissions before publishing
• Category browsing across coding, testing, security, and more

Pros

• Security scanning catches the skills that could quietly cause harmful or wasteful agent behavior
• Large indexed catalog with continuous updates
• Free with a simple CLI install

Cons

• Security scanning isn't the same as benchmarking, so no evidence a skill reduces tokens
• Self-reported index numbers change continuously and aren't independently verified

Visit Skills Directory (skillsdirectory.com)

Conclusion

Token waste in 2026 mostly comes from three places: agents building the wrong thing, context windows that bloat over long sessions, and skills that load without earning their keep. The skills on this list attack those problems differently. caveman compresses output directly, GSD Core runs work in fresh windows, grill-me kills wrong-direction builds early, and the Karpathy rules push agents toward surgical changes. But before you install any of them, it's worth checking whether they actually deliver, because a skill that claims savings and doesn't is just more overhead. That's where SkillzBench comes in. It's free, it security-scans every versioned skill, and it benchmarks each one in a sandbox to publish the measured Lift, so you can see what a skill really does before it touches your context. The free MCP connector and per-agent loadouts then let you equip only what each agent needs. Skills, proven. Browse the directory and current benchmark results at skillzbench.com, and start with a leaner, verified loadout today.

Frequently asked questions

What makes SkillzBench different?

SkillzBench doesn't just list skills, it tests them. Every skill version is pinned to a commit SHA, security-scanned, license-checked, and then benchmarked in a sandbox with the same model, with and without the skill. The published Lift in percentage points shows whether a skill actually improves results. It also gives you a free MCP server to connect any agent to your chosen skills, plus per-agent loadouts so each agent only sees what you toggled on. Everything on the platform is free, and advertising revenue never influences scores or listings.

Do Claude skills really reduce token usage?

Some do, and some don't. Skills like caveman exist specifically to compress output, and skills like grill-me or the Karpathy guidelines reduce waste by preventing agents from building the wrong thing. Claude's own skill design helps too: skills load only when a task matches their description, so a skill costs almost nothing until it's needed. But skills that orchestrate heavily, like Superpowers or GSD Core, can add their own token overhead. The only way to know for your workload is to measure, which is what SkillzBench's sandbox benchmarks do.

Are agent skills safe to install?

Not automatically. Published research found roughly 26% of skills carry at least one vulnerability, and Skills Directory reports that 36% of skills in the wild have security flaws, including prompt injection and credential theft. Before installing, check whether the skill has been security-scanned, whether it's pinned to a specific commit, and whether anyone has independently tested it. SkillzBench runs security scans and publishes a 0 to 100 trust score for every listed skill.

How many skills should I install?

Fewer than you think. Every loaded skill costs a little context, and consultants' guides recommend starting lean, often with just a handful covering your actual daily work. Per-agent loadouts help here: rather than giving every agent every skill, toggle on only what that agent needs. That keeps the agent's visible skill set small and its context clean.

Is there a cost to using these skills and platforms?

The skills on this list are free and open source, and the discovery platforms (skills.sh, Skills Directory, SkillzBench) are free to use. The main cost is tokens: your Claude plan or API usage still covers whatever the agent consumes, which is exactly why choosing skills that reduce waste matters.

Do these skills work outside Claude Code?

Mostly yes. The SKILL.md format works across Claude Code, Codex, Cursor, Gemini CLI, and other agents that adopted the standard. Superpowers, GSD Core, and Matt Pocock's skills all document multi-agent support. SkillzBench's free MCP connector is also agent-agnostic, so you can connect whichever agent you run to the skills you choose.