`claude-api` skill dumps entire ~784KB SKILL.md into context for a simple usage-limit question, burning huge Pro-plan usage

Description

Asking a plain informational question that happens to mention a model name (e.g. "does Fable count towards the 5-hour usage limit, and how does its consumption rate compare to Sonnet and Opus?") causes Claude Code to invoke the bundled claude-api skill, which dumps its entire SKILL.md — ~784KB / ~195K+ tokens — into the conversation as a single tool result. This happens even though the question has nothing to do with writing code; it's a plan/usage/pricing question, not an SDK implementation request.

The claude-api skill's trigger condition fires on almost any mention of Claude/Anthropic/model names ("the prompt names Claude/Anthropic in any form... claude-*... OR the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory"), so a short usage-policy question is enough to load the full skill body — including unrelated sections on streaming, tool use, MCP, prompt caching, code-execution sandboxes, etc. — instead of just the (much smaller) relevant excerpt.

Impact

On a Pro plan, a single question triggered one Skill call that alone consumed on the order of 200K tokens of context, on top of normal conversation overhead. This happened identically three separate times (I re-asked variants of the same question across sessions), each one ballooning to ~900KB and burning a large chunk of the 5-hour usage window for what should have been a one-line factual answer. For a Pro-tier user this is a disproportionate amount of the available usage budget spent on a single skill-load side effect.

Steps to reproduce

1. On Claude Code 2.1.215, start a fresh session.
2. Ask: does fable count towards 5-hour usage? if so, how is the consumption compared to sonnet and opus
3. Observe that Claude invokes Skill(skill: "claude-api", ...).
4. Inspect the session transcript (~/.claude/projects/<project>/<session-id>.jsonl) — the tool_result for that Skill call is ~784KB of text (the full claude-api SKILL.md), for a question that required no code generation at all.

Expected behavior

A simple factual/pricing/usage-limit question shouldn't require loading a ~200K-token, code-generation-oriented skill document in full. At minimum:

• The claude-api skill's trigger condition should be narrower (e.g. distinguish "asking about pricing/limits" from "asking to write code against the API"), or
• The skill should support loading only the relevant section instead of the entire file, or
• Large bundled skills like this should be chunked/paginated rather than injected as one monolithic tool result.

Environment

• Claude Code version: 2.1.215
• OS: Linux (WSL2), Ubuntu on Windows
• Plan: Pro
• Model in use when triggered: Sonnet 5 (default)

Additional context

This is filed as a genuine "please look into this" bug report — I'm a Pro-tier user and this ate a large amount of my 5-hour usage window on a single low-value exchange, three times in a row. If there's any consideration for a usage reset or a plan adjustment given this was a product-side inefficiency rather than something I did, I'd genuinely appreciate it — but the main ask is fixing the oversized skill dump.