cheaper-skill-for-claude

Cheaper Skill For Claude

Purpose

Use this skill to run Claude Code workflows in a cache-friendly way without changing Claude Code source code. The main goal is to make repeated input cheaper by improving prompt-cache reuse, not by hiding context, replaying old answers, or changing model outputs.

This skill is Claude-specific. Keep the outer project name broad, but treat this skill as the Claude Code adapter because each agent framework exposes a different prompt assembly surface.

Codex may use this skill to help operate the workflow, but the experiment target is Claude Code. In the current setup, Claude Code is routed to a MiMo model such as mimo-v2.5-pro. Keep those roles separate in reports:

• Codex: development assistant and skill runner
• Claude Code: coding-agent harness under study
• MiMo: backend model/provider route
• make-agents-cheaper: audit/eval instrumentation

Core Rule

Structure the prompt so stable components come first and dynamic components come later.

For Claude Code, prefer the official CLI flag:

claude --exclude-dynamic-system-prompt-sections

This moves per-machine dynamic sections such as cwd, env info, memory paths, and git status out of the early system prompt and into the first user message. The dynamic information is still present, but the early prefix becomes more stable across runs and users.

The flag only applies to Claude Code's default system prompt. If a command uses --system-prompt, do not assume this flag is active.

Quick Start

Interactive run:

claude --model mimo-v2.5-pro --exclude-dynamic-system-prompt-sections

Print-mode measurement run:

claude -p \
  --model mimo-v2.5-pro \
  --output-format json \
  --no-session-persistence \
  --exclude-dynamic-system-prompt-sections \
  "$PROMPT"

If the provider/model is not MiMo, keep the same cache-friendly structure but update the model name and token-field parser.

Guardrails

• Keep provider, model, reasoning effort, tools, MCP servers, hooks, and working directory stable across compared runs.
• Do not modify Claude Code source code for the basic skill workflow.
• Do not remove important context just to reduce visible prompt size.
• Do not claim savings from a single cold run; cache-shape changes usually require warm-up.
• Do not print API keys, auth tokens, or full raw request bodies that may contain secrets.
• Do not call it universal until a separate adapter exists for that agent.