What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The claimed 92% reduction—more than 24,800 tokens per session—is an author-reported result, not a figure independently verified by the available documentation. The underlying technique is progressive disclosure: keep a skill’s full instructions out of the always-present prompt and load them only when a task needs them. That can reduce baseline context, but metadata still takes space, and a skill’s instructions consume context once loaded.
What the 92% token claim does—and does not—show
The title’s figures describe the author’s reported experience across more than 200 Claude Code skills and Antigravity agents. Without the original before-and-after session data and a stated measurement method, they should not be treated as a reproducible result or a prediction of what another setup will save.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
To evaluate the claim, a useful report would specify the baseline and optimized configurations, the tasks and session boundaries, model and version, test date, which token categories were counted (input, output, cached, or total usage), and whether answer quality was comparable. The cited product documentation does not supply those details for this setup.
Why loading skills only when needed can reduce baseline context
Anthropic describes Agent Skills as filesystem-based bundles of instructions and resources. Its documented progressive-disclosure approach makes skill metadata available first, then reads a skill’s full instructions when relevant; additional resources can be accessed as needed. This avoids putting every full skill body into every request when the implementation follows that pattern. See Anthropic’s Agent Skills overview.
#1 Best Overall
This is a distinction between baseline context—what is present before a task invokes a skill—and loaded context, which grows when task-specific instructions or resources are read. A large library can therefore have a smaller baseline than one monolithic prompt containing every instruction, but it is not cost-free: metadata is still present, and loaded instructions compete with the conversation and other context. Anthropic’s authoring guidance puts the principle plainly: “The context window is a public good.” Its advice is to keep skills concise and avoid loading unnecessary material. See Anthropic’s skill-authoring best practices.
Progressive disclosure versus one large prompt
| Approach | What is present at the start | What happens during a task | Main consideration |
|---|---|---|---|
| Monolithic prompt | All bundled instructions are included in the baseline. | The instructions are already present, whether or not a given task needs them. | Simple to inspect, but unrelated material occupies context. |
| Progressive disclosure | Skill metadata is available; full skill instructions are deferred. | Relevant instructions and resources are loaded as needed. | Can reduce baseline context, but metadata remains and invoked instructions use context. Selection depends on useful descriptions and relevance. |
Google’s 2026 ADK guide gives an illustrative comparison of roughly 1,000 tokens for L1 metadata versus 10,000 tokens for a monolithic prompt with ten skills—roughly a 90% reduction in baseline context in that example. It is an example about Google ADK, not a Claude Code or Antigravity benchmark and not verification of the 92% claim above. See Google’s ADK guide.
What the evidence establishes about Claude Code and Antigravity
Claude Code
Anthropic’s Agent Skills documentation supports explaining staged skill loading in its documented system. That does not establish the token behavior or savings of every custom Claude Code setup. Claude Code usage and metering also depend on sign-in and billing configuration, so a reported token count should identify the usage view and categories behind it.
Antigravity
The available Google source is a search result dated May 19, 2026 announcing a transition from Gemini CLI to Antigravity CLI and describing it as an agent-first platform. It does not document the loading semantics of the author’s setup or confirm that the same 200-plus-skill arrangement works in Antigravity. See the Google Developers Blog search result. Claude’s documented behavior should not be assumed to apply to Antigravity.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsHow to measure token use in Claude Code
For API usage, Anthropic Support documents the /cost command for inspecting a session’s token and dollar usage. The available help material does not prescribe a full experiment, so define the comparison carefully rather than reading a single session as proof of a percentage reduction. See Anthropic Support’s Claude Code usage guide.
- Set the measurement boundary. Decide what counts as a session and whether you are comparing input, output, cached tokens, or total usage. Record the model/version, settings, and date.
- Record a baseline. Run a defined set of representative tasks with the existing configuration and use
/costto inspect API session usage. - Change the context design. Move task-specific material out of always-present instructions where the implementation supports deferred loading. Keep descriptions clear enough for relevant skills to be selected.
- Repeat comparable tasks. Use the same task set, session boundaries, model, and settings as far as possible. Compare like token categories and report the number of sessions, not just one favorable run.
- Check output quality. Compare whether the results still meet the same requirements. Token savings alone do not show that the task was completed equally well.
One context cost that is easy to miss
Adding a file with @ can expand the prompt beyond that file: Anthropic Support warns that it includes the file and its CLAUDE.md tree in context. Check what is being pulled in when a request unexpectedly uses more context than intended. Keep references targeted and avoid injecting large files or instruction trees when the task does not need them.
Quick Recap
Rank #4
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.
