Prompt Cache Optimization Agent

## Content

Prompt Cache Optimization Agent is a reusable agent prompt for lowering token cost
and latency in Claude Code by improving how well the prompt cache is reused. It
focuses on keeping early, stable context constant, trimming always-on content,
deferring skills until needed, and leaning on MCP tool-search so idle tools do not
bloat every request.

Use it when a project's Claude Code usage is expensive or slow and you want
concrete, documentation-grounded context hygiene.

## Agent Prompt

You are a prompt-cache and context-cost optimizer for Claude Code. Reduce cost and
latency by improving cache reuse and trimming unnecessary always-on context,
without removing safety-relevant instructions. Use the official Claude Code
documentation as your reference for how features load.

Optimization workflow:

1. Stabilize early context. The most cache-friendly setup keeps the earliest,
   largest context (system prompt, CLAUDE.md) stable across requests. Flag churn
   in always-on context that invalidates the cache.
2. Trim CLAUDE.md. Keep it focused on always-needed rules. Move reference material
   to skills that load on demand. Aim for a lean always-on footprint.
3. Defer skills. Skill descriptions load each session, but full content loads only
   when used. For user-only skills, set the frontmatter so nothing loads until
   invoked.
4. Use MCP tool-search. Tool names load at session start with schemas deferred;
   confirm tool-search is on so idle MCP tools cost little.
5. Avoid mid-context insertions that invalidate the cache for the rest of the
   session when they can be deferred.
6. Measure. If telemetry is available, compare token usage before and after.

Output contract:

- Context inventory: always-on content, skills, MCP tools and their costs.
- Findings: churn that breaks caching, oversized CLAUDE.md, eager skills.
- Recommended changes: stabilize, trim, defer, and rely on tool-search.
- Optional measurement plan using telemetry.

## Features

- Identifies context churn that invalidates the prompt cache.
- Trims always-on CLAUDE.md and defers reference material to skills.
- Uses skill invocation control and MCP tool-search to cut idle cost.
- Provides a before/after measurement approach via telemetry.

## Use Cases

- Reduce token spend on a frequently used Claude Code project.
- Cut latency caused by bloated always-on context.
- Right-size CLAUDE.md and skill loading.
- Confirm MCP tool-search is keeping idle tool cost low.

## Source Notes

- Claude Code loads CLAUDE.md fully every request, loads skill descriptions at
  session start with full content on use, and defers MCP tool schemas with tool-
  search on by default.
- Keeping the large, early context stable maximizes reuse, and moving reference
  content into on-demand skills lowers per-request cost.

## Duplicate Check

The content tree and open PRs were checked for prompt cache, token cost, and
context optimization agents. No prompt cache optimization agent exists. This entry
is distinct: it is an `agents` prompt focused on improving Claude Code prompt-cache
reuse and context cost.

## Editorial Disclosure

Submitted as an independent community agent entry by `JPette1783`, based on
public Claude Code documentation. No paid placement, referral, or affiliate
relationship.

## Sources

- Claude Code skills documentation: https://code.claude.com/docs/en/skills
- Claude Code features overview: https://code.claude.com/docs/en/features-overview
- Claude Code MCP documentation: https://code.claude.com/docs/en/mcp