Recommended Free Tools
The claimed 92% reduction—more than 24,800 tokens per session—is an author-reported result, not a publicly verified benchmark. The defensible idea behind it is progressive disclosure: keep a compact index of reusable skills available, then load a skill’s full instructions and resources only when a task needs them. That can reduce always-present context, but it does not make a large skill library free: metadata still occupies context, and instructions consume tokens when loaded.
What the 92% claim means—and what it does not establish
The title describes a personal result across an author-reported collection of more than 200 Claude Code skills and Antigravity agents. The available evidence does not include session logs, a benchmark method, or a verified inventory, so the 92% and 24,800-plus-tokens-per-session figures should be treated as the author’s report rather than a result others can reproduce.
As an Amazon Associate I earn from qualifying purchases.
To evaluate or repeat that result, a comparison would need to disclose the baseline and optimized configurations, matched tasks and session boundaries, model and settings, test date, and the token categories counted—such as input, output, cached tokens, or total usage. It should also show that the compared sessions produced work of comparable quality. Without those details, the percentage cannot be independently interpreted.
How progressive disclosure can reduce baseline context
Anthropic describes Agent Skills as filesystem-based bundles of instructions and resources. Its documented pattern makes skill metadata available first, loads full instructions when a skill is relevant, and allows additional resources to be accessed as needed. This is progressive disclosure: an agent can have a broad library available without inserting every full skill into every task’s initial context. Anthropic’s Agent Skills overview explains the model.
The distinction is between baseline context, present at the start of an interaction, and task-specific context, added when the agent invokes a relevant skill or reads a resource. A monolithic prompt that includes all skill instructions pays for that material up front. A staged setup can defer full skill bodies, but a selected skill still adds its instructions when used.
What remains in context
- Skill names and descriptions or other metadata may be preloaded, so the library can still have a baseline context cost.
- Once a skill is loaded, its instructions compete with conversation history and other context.
- Referenced files and resources add context when read; a large library does not imply that all its content is costless or automatically excluded.
- Selection quality matters: useful descriptions help an agent identify relevant skills, while overly broad instructions or unnecessary loads can erode the benefit.
Anthropic’s authoring guidance recommends keeping skills concise and notes that loaded skill text competes with the rest of the context. It captures the underlying constraint in one sentence: “The context window is a public good.” Anthropic’s skill-authoring best practices provide the guidance.
Rank #2
How this compares with Google’s published example
Google’s 2026 ADK walkthrough illustrates a separate example: roughly 1,000 tokens for Level 1 metadata versus 10,000 tokens for a monolithic prompt containing ten skills, or about a 90% reduction in that example’s baseline context. The comparison is about Google’s illustrated ADK setup; it is not a test of Claude Code, Antigravity, or the author’s 200-plus-skill arrangement. Google’s ADK guide describes the example.
| Approach | What enters context initially | What happens when a task needs a skill | What the evidence supports |
|---|---|---|---|
| Monolithic prompt | All bundled skill instructions are present in the baseline. | The task can use instructions already included, but unrelated material also occupies context. | Architectural comparison; no matched performance benchmark is established here. |
| Progressive disclosure | Metadata or descriptions are available first; full bodies are deferred. | Relevant instructions and any needed resources are loaded, adding context at that point. | Anthropic documents this pattern for Agent Skills; Google’s token figures apply only to its separate ADK illustration. |
How to measure token use in Claude Code
First define what you want to compare. A reduction in always-present prompt context is not necessarily the same as a reduction in total billed usage: a task that loads a skill later still uses tokens for that material, and output length and other input can also affect session totals.
Rank #3
- Choose a consistent workload. Run the same set of tasks with the same model and comparable settings, using the same session boundaries. Record the date and the configuration for each run.
- Separate the configurations. Keep the baseline and optimized skill-loading arrangements explicit. Note which instructions are always present and which are loaded only when invoked.
- Record token categories. State whether the comparison includes input, output, cached tokens, or all usage. Do not compare a baseline-context count with a total-session count as though they were the same measure.
- Inspect supported usage information. Anthropic’s help page documents
/costfor inspecting session token and dollar usage when using API billing. Sign-in and metering setup affect the usage experience, so check the current instructions for your account and setup. Anthropic’s Claude Code usage and limits help page describes the details. - Check task results as well as usage. Keep outputs or apply a consistent quality check. Lower token use alone does not show that the two configurations did equivalent work.
Check file injection and tool-specific assumptions
Claude Code can add more context than the file named in a prompt: Anthropic’s help material warns that using @ to inject a file includes that file and its CLAUDE.md tree in context. If token use rises unexpectedly, inspect injected files and project instructions as well as skills.
Do not assume Claude Code’s loading behavior describes Antigravity. The available Google material identifies a transition announcement describing Antigravity CLI as an agent-first platform, but does not establish that the author’s particular skill-loading arrangement or token results are an official Antigravity capability. Google’s announcement search result supplies only that adjacent product context; implementation details and the claimed cross-tool savings remain unverified.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →
