Skill Tokens: Why Every Skill Costs You Context

The Hidden Bill: How Skill Bloat Inflates Your Token Costs

Here's the uncomfortable math: the skills you install to make your agent smarter often make your bill bigger. Agent skills were designed to save tokens by encapsulating reusable knowledge, yet a large-scale study of 55,315 public GitHub skills found they frequently increase token consumption as currently authored.

Start with the baseline. Before you type a single word, a typical agent session loads 20,000–30,000 tokens of system prompt, config files, memory, and tool schemas. One developer documented a "hi" prompt consuming roughly 31,000 tokens, according to token-efficiency research on Claude Code usage. Every skill you add rides on top of that hidden bill.

The skill content itself compounds the problem. The same study found only 38.5% of skill body content is actionable rules — 40.7% is background material the model rarely needs. Meanwhile, 26.4% of public skills lack descriptions entirely, meaning they consume router tokens on every invocation while never matching a relevant request.

Multiply that across a team's usage and the costs get real:

• A 10,000-token skill costs $0.03–$0.15 per invocation at current API pricing.
• Teams with frequent agent usage can spend hundreds of dollars monthly on skill content alone.
• Each installed skill adds 30–100 tokens at session start, before it ever fires.
• MCP servers add 10,000–20,000 tokens per server, per session.

There's also a performance cost that doesn't show up on any invoice. Anthropic's context engineering guidance describes LLMs as operating on a finite "attention budget" — every new token depletes it, stretching the model's ability to capture relationships across its context. Filler isn't just expensive; it's distracting.

The proof is in the trimming. Counterintuitively, the SkillReducer study found that removing non-essential skill content improved functional quality by 2.8% over the originals — a less-is-more effect where cutting filler reduces distraction. A practitioner who audited their setup cut 405KB of overhead down to 232KB, a 44% reduction, with what they described as the same capabilities.

This is why the emerging discipline is to treat every skill as a line item with a measurable context cost. It's the gap SkillzBench is being built to close: a free directory and MCP connector that pins, scans, and benchmarks skills, publishing the Lift in percentage points so you can see whether a skill earns its token bill before you pay it. Skills, proven.

Attention Is Not Free: Why More Context Makes Your Agent Worse

Every token you feed a model is a withdrawal from a finite account. Anthropic calls it the attention budget: transformers build pairwise relationships between every token in the window, so as context grows, the model's ability to capture those relationships gets stretched thin.

This produces "context rot" — the well-documented phenomenon where a model's ability to accurately recall information decreases as the window fills. As one Anthropic engineer put it, quoted in recent analysis: "The context is not full, it is poisoned. The model is not broken. The context is."

The empirical evidence for a less-is-more effect is striking. A large-scale study of 55,315 public skills found that trimming non-essential content improved functional quality by 2.8% over the originals — removing background material reduced distraction rather than losing information. Only 38.5% of typical skill body content is actionable core rules; 40.7% is background and 12.9% is examples.

SkillsBench, the first peer-reviewed academic benchmark for agent skills, corroborates this from the performance side. Its findings, reported in detail here, show a clear curve:

• 2–3 focused skills: +18.6 percentage points over the no-skill baseline
• 4+ skills: +5.9 points — returns collapse as context fills
• Monolithic "comprehensive" skills: -2.9 points — actively harmful
• Self-generated skills: 8.1–11.5 points *below* the no-skill baseline across Claude Code, Codex, and Gemini CLI

That last number deserves attention. When agents write their own skills, performance drops below using no skills at all. As the benchmark authors conclude, models cannot reliably author the procedural knowledge they benefit from consuming.

The bottleneck isn't model intelligence — it's context. A smaller model with the right three skills outperforms a frontier model with none: Claude Haiku 4.5 with curated skills scored 27.7% versus Claude Opus 4.5 without them at 22.0% in the same benchmark.

This is the engineering problem SkillzBench is being built to solve: pinning every skill to a commit SHA, security-scanning it, and benchmarking it with and without the skill so you can see the Lift before you spend the context. The directory, MCP connector, and benchmarks are all free — join the waitlist to get notified at launch.

The Fix: Fewer, Smaller, Measured Skills

The evidence points to a clear discipline: treat skills like any other dependency, and make each one earn its place. Start with an audit. A practitioner who cut global skills from 20 to 6 reduced total context overhead by 44% with the same capabilities, simply by moving rarely used skills out of the always-loaded set. That aligns with benchmark data showing 2–3 focused skills per task deliver the strongest gains, while monolithic bundles actively hurt performance.

The mechanics matter as much as the count. Restructuring monolithic skills into a tiered, progressive-disclosure architecture — a compact core loaded always, background modules fetched on demand — achieved 48% mean description compression and 39% body compression in one study, while functional quality actually improved by 2.8% over the originals. Fewer tokens, better results.

A workable sequence looks like this:

• Audit every installed skill and cut to 1–3 per task, keeping only those used in most projects.
• Split each remaining skill into an actionable core and on-demand reference material.
• Gate adoption behind a with-versus-without evaluation — keep what measurably helps, discard what does not.
• Re-check after upstream changes, since a skill that helped last month may not help today.

That last point is where most teams stall. Self-generated skills, for instance, landed 8.1 to 11.5 points below the no-skill baseline across Claude Code, Codex, and Gemini CLI, which means an evaluation gate is not optional. This is the gap SkillzBench is being built to close: a free directory and MCP connector that pins every version to a commit SHA, scans it, and benchmarks it with and without the skill, publishing the Lift in percentage points so nobody has to guess whether a skill earns its context cost.

Make It Routine: How to Prove a Skill Earns Its Context

Proving a skill earns its context is not a one-time judgment call — it is a routine you run on every skill you adopt. Treat each skill as a line item with two numbers attached: what it costs in tokens, and what it returns in Lift.

The cost side is measurable. A 10,000-token skill runs $0.03–$0.15 per invocation at current API pricing, and teams with frequent agent usage can spend hundreds of dollars monthly on skill content alone, according to a large-scale study of 55,315 public GitHub skills. That same research found 26.4% of public skills lack descriptions entirely and 44.1% are missing or under 20 tokens — consuming router tokens on every call while being impossible to match to a request.

The Lift side requires a benchmark, not a vibe. The discipline is simple: run the same model on the same task with and without the skill, then keep only what measurably helps. SkillsBench, the first peer-reviewed academic benchmark on agent skills, found 2–3 focused skills yielded +18.6 percentage points while monolithic skills hurt performance by -2.9 points — evidence that the number, not just the content, determines whether a skill pays off.

A repeatable workflow looks like this:

• Cost it: estimate tokens per invocation and multiply by real call volume.
• Benchmark it: test with and without the skill on identical tasks.
• Demand a Lift number in percentage points before installing.
• Cut or restructure anything that adds tokens without adding Lift.
• Re-verify when upstream changes land — a new model version can invalidate an old result.

That last step is the one most teams skip. Skills are versioned artifacts, and a skill that helped last quarter may not help after an upstream update. SkillzBench is being built as a free, independent directory and MCP connector to make this routine practical: pinning every skill version to a commit SHA, security-scanning it, and publishing Lift in percentage points alongside a separate trust score — so you never pay the context cost of a skill that doesn't help.