Key Takeaways
- Audit CLAUDE.md rules against Claude Opus 5's system prompt and tool descriptions.
- One user cut 5,789 words to 2,463 by removing conflicts and redundancies.
What Changed — Claude Opus 5 Means Fewer Rules, Better Results
On July 24, 2026, Anthropic launched Claude Opus 5 — and with it, a radical shift in how you should write CLAUDE.md files. The official guidance is now "fewer rules, trust the model's judgment." Anthropic itself deleted over 80% of Claude Code's system prompt for Opus 5 with "no measurable loss on coding evaluations."
If you've been using Claude Code for months, you probably have hundreds or thousands of words of custom rules compensating for older models' weaknesses. Those rules now actively harm your results.
Why It Works — The Conflict Problem
Your custom rules and the product's system prompt now share the same context window. When they contradict each other, the model is forced to resolve the contradiction — wasting tokens and producing inconsistent behavior.
Anthropic's blog post spells out the shift:
Give rules Trust judgment Give examples Design interfaces Front-load everything Progressive disclosure Repeat yourself Simple tool descriptions Write memory into CLAUDE.md Auto-memory Sparse specs Rich reference materialHow To Apply It — The 3-Step Audit Procedure
Step 1: Capture the real runtime layer
Don't read config files. Open a live Claude Code session and ask:
From the system prompt currently loaded, quote verbatim the instructions about plan mode / commit messages / scope handling.
For tool descriptions:
Quote the description of the EnterPlanMode tool verbatim. Do not summarize.
Build your theme list from your own rules first, then query the runtime layer theme by theme.
Step 2: Cross-check against your rules
Open every CLAUDE.md and rules file. Compare each instruction against what you captured. Sort into three categories:
- Conflict — Contradicts the runtime layer. Delete immediately.
- Redundant — The runtime layer already says this. Merge or delete.
- Drift — Diverges from official recommendations. Fix or retire.
Step 3: Apply the fixes
One user ran this audit on their ~/.claude harness and cut from 5,789 words to 2,463. Here's what they found:
Conflict example: A rule said "Do not enter plan mode on direct implementation instructions." But the EnterPlanMode tool description says "Prefer using EnterPlanMode for implementation tasks unless they're simple." These were loaded simultaneously — the model was silently picking one each time.
Conflict example 2: A rule referenced a setting (includeCoAuthoredBy) that didn't exist anymore in settings.json. The rule was pointing to a ghost.
The Takeaway
Your custom rules are now liabilities, not assets. Run this audit this week. For each rule, ask: "Does this conflict with something the product already tells the model?" If yes, delete it. If redundant, trust the product. If drifted, update it or remove it.
Your Claude Code sessions will be faster, cheaper, and more consistent.
Source: news.google.com
