Your CLAUDE.md Is Now Technical Debt: What Claude Opus 5 Wants You to Delete

Claude Opus 5 shipped on Friday: $5 and $25 per million tokens, unchanged from Opus 4.8, within half a percent of Fable 5 on CursorBench at half the price, a 1M token context window as both default and maximum. The number that should worry your config file is a verified 30.16% on ARC-AGI 3, against 7.8% for the next-best model, on a benchmark built specifically so that memorisation doesn’t help.

The interesting document isn’t the announcement. It’s the migration guide, whose checklist runs through the usual chores, update the model string, revisit max_tokens because thinking is now on by default, and then, sitting among them as if it were the same kind of task:

Remove verification/self-check instructions.

Software migrations don’t usually ask you to delete English.

The Instructions That Turned Against You

Opus 5 verifies its own work and catches its own mistakes unprompted, so the prompting guide says instructions telling it to do those things “compound with the model’s own behavior and add cost without improving results.” Two ways your existing rules now hurt:

• They bill. On a weaker model a redundant instruction was free: it got ignored. On a model that obeys, “double-check your answer” buys a full extra pass at $25 per million output tokens, every turn, to redo finished work. One HN commenter benchmarked the same build at 470k tokens and $1.29 on Opus 4.8 against 179k and $0.33 on Opus 5. A stale instruction spends that efficiency straight back.
• They mean something new. Literal following turns your hedges into specifications.

If your review prompt says “only report high-severity issues” or “be conservative,” the model may follow that instruction literally and report less.

— Anthropic, Prompting Claude Opus 5

You wrote “be conservative” to talk down a model that fabricated findings. On Opus 5, which the same guide describes as finding real bugs at a high rate with few false positives, it suppresses true positives instead, and the only symptom is bugs you never hear about. The model didn’t get better at following your good instructions. It got better at following your instructions. When I had Grok review its own integration, eight of its ten findings were noise, and the obvious reflex is to tell the reviewer to be conservative; aim that same instruction at a precise model and you have built a filter that discards signal.

Anthropic Deleted Four Fifths of Its Own

Thariq from the Claude Code team published the postmortem the morning after launch:

We removed over 80% of Claude Code’s system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations.

— Thariq, Claude Code team, Anthropic

Their diagnosis was over-constraining. Reading their own transcripts they found the system prompt, skills and CLAUDE.md files contradicting each other inside a single request: “leave documentation as appropriate” against “DO NOT add comments.” Their comment policy is the cleanest before and after. Old: “default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks.” New, currently shipping: “Write code that reads like the surrounding code: match its comment density, naming, and idiom.” A prohibition became a judgement call, because the model can now make it, and the ARC-AGI 3 jump is the evidence for why. Two of their other findings invert advice I have given repeatedly: examples now constrain rather than clarify, and the everything-in-one-file CLAUDE.md is named a myth. Their own verification instructions weren’t deleted, just moved into a skill that loads when needed.

So I Audited Mine

Your package.json pins versions. Your CLAUDE.md is unversioned prose, read every turn, evaluated by an interpreter that gets silently replaced every eight weeks. So I ran the checklist across every CLAUDE.md and AGENTS.md under my projects directory: 74 files, 14,179 lines, personal and work. Grepping for the deletion target returned 45 lines mentioning verification, and none of them were the thing Anthropic told me to delete.

• Domain vocabulary. JWT verification, HMAC signatures, session middleware. The word is there because the codebase does that thing.
• External gates. “Always run flutter analyze before committing.” Across the fleet, 190 lines pair a run-or-verify verb with a named tool.
• Genuine self-check imperatives: zero, in all 74 files.

The checklist collapses two meanings of one word. Anthropic’s “verification” is the model re-reading its own work, now native. A CLAUDE.md’s is run this specific command, and Opus 5 cannot know your linter exists or whether CI went green. Grep-and-delete would have stripped my lint gates and left the real redundancy untouched. The actual debt was the inverse: seventeen of 74 files say anything about length, and Opus 5 writes longer by default. The stale rule wasn’t one I needed to delete. It was one I had never written, because the model I wrote the file for didn’t need it.

What This Doesn’t Fix

• Anthropic’s evals are not your monorepo. “Removing it costs no quality” is measured on their distribution. Delete in small batches and watch, the same way you would with a caching change.
• The advice is credible because it costs them money. A vendor billing per token telling you to delete token-burning instructions is a claim against interest. Notice the same guide steers demanding work toward xhigh and max, which isn’t.
• Your file won’t get shorter so much as move. No verification rule survived my grep to be deleted, and the file still owed additions: conciseness, narration cadence, delegation caps. Then I ran /doctor, Claude Code’s new rightsizing command, against this blog’s own 347-line instruction file. It came out at 104, with 197 lines relocated into a skill that loads only when I’m writing a post: about 2,800 tokens back every session, four fifths of it relocation rather than deletion. Same move Anthropic made with verification, and it falls due again in eight weeks.

Anthropic rewrote its prompts for the model it built, deleted four fifths of them, and handed you the delta. Yours is still sitting in the repo root, unversioned, unreviewed, and load-bearing on every turn you have run since Friday.