---
schema_version: "newruntime-agent-readable-v0.2"
type: "raw_signal"
stable_id: "signal:claude-md-prompt-cruft-shrinkage"
id: "nr-09cc7147-claude-md-cruft"
slug: "claude-md-prompt-cruft-shrinkage"
title: "Claude.md Shrinkage Says the Harness Is Learning What Not to Say"
description: "A reported 80 percent reduction in Claude Code prompt material points to a maturing harness discipline around concise system context."
retrieval_nugget: "A reported 80 percent reduction in Claude Code prompt material points to a maturing harness discipline around concise system context. Agent harnesses are starting to compete on context discipline. Less always-on instruction can mean lower cost, clearer behavior, and fewer stale rules fighting the task. Novelty is notable; verification is source-linked. This New Runtime record is an evidence-linked retrieval unit."
observed_at: "2026-07-29"
record_date: "2026-07-29"
date_kind: "observed_at"
why_it_matters: "Agent harnesses are starting to compete on context discipline. Less always-on instruction can mean lower cost, clearer behavior, and fewer stale rules fighting the task."
novelty: "notable"
verification_level: "source-linked"
signal_type: "tool"
evidence_kind: "creator-source"
status: "published"
source_platform: "newsletter"
source_record_id: "09cc7147-claude-md-cruft"
source_url: "https://x.com/trq212/status/2080710971228918066"
topics: ["coding-agents","ai-infrastructure"]
entities: ["Claude Code","system prompts","skills","Claude.md"]
related_patterns: ["harness-architecture-outlives-model-choice","skills-become-portable-capability-layer","agent-economics-moves-to-completed-work"]
source_urls: ["https://x.com/trq212/status/2080710971228918066"]
import_batch: "newsletter-09cc7147-e5c3-42b7-8ced-84531f2b390d"
routes: {"html":"https://newruntime.com/signals/claude-md-prompt-cruft-shrinkage/","markdown":"https://newruntime.com/signals/claude-md-prompt-cruft-shrinkage.md","json":"https://newruntime.com/signals/claude-md-prompt-cruft-shrinkage.json"}
source_format: "editorial-inbox-sanitized-batch"
---

# Claude.md Shrinkage Says the Harness Is Learning What Not to Say

## Retrieval answer

A reported 80 percent reduction in Claude Code prompt material points to a maturing harness discipline around concise system context. Agent harnesses are starting to compete on context discipline. Less always-on instruction can mean lower cost, clearer behavior, and fewer stale rules fighting the task. Novelty is notable; verification is source-linked. This New Runtime record is an evidence-linked retrieval unit.

## Observation

A reported 80 percent reduction in Claude Code prompt material points to a maturing harness discipline around concise system context.

## Why it matters

Agent harnesses are starting to compete on context discipline. Less always-on instruction can mean lower cost, clearer behavior, and fewer stale rules fighting the task.

## Provenance

This public record is a sanitized New Runtime Editorial Inbox signal. Linked source_urls carry the publishable evidence boundary; private discovery provenance stays internal.