Anthropic Says Delete 80% of Your Prompt Files, and Three Independent Sources Agree

For two years the technique was accumulation. Longer system prompts, longer CLAUDE.md, more numbered do/don't lists, more "always verify your work" imperatives. Anthropic's context-engineering guidance for Claude 5 models inverts it, with an 80% deletion figure attached. The stated reason isn't that the rules were wrong. It's that they contradicted each other inside a single request and suppressed exploration in cases where exceptions were correct.

Replacement pattern: state the goal, let the model judge, load context progressively.

I went looking for whether this holds up outside Anthropic's own blog, and three separate lines of evidence converge.

First, run /doctor. It's a nine-point read-only audit of your Claude Code setup: install health, duplicate installs, broken settings, colliding agent definitions, dead weight, CLAUDE.md redundancy, lazy-loading opportunities, context budget, permissions. It reports before it changes anything. In one live run it found 113 personal skills burning roughly 10,000 tokens of context every single session just to be listed, about a quarter of which had never been used. That's not a prompt-quality problem, that's paying rent on an empty apartment.

Second, paddo.dev audited 74 CLAUDE.md and skill files against the new guidance and found zero genuine self-check imperatives: most "verification" mentions turned out to be lint gates or domain terms. The delete list is instructions like "double-check your answer," which Opus 5 does natively and which you're billed for twice. The keep list is domain vocabulary, external tool commands, and repository facts the model can't deduce from file structure. Their structural fix is the part I'm stealing: split CLAUDE.md into model-behavior rules and repository facts, and put a review date on the first section only. Model capabilities shift roughly every eight weeks, so today's guardrail is next quarter's contradiction. That makes the cleanup repeatable instead of a one-time purge.

Third, and this is the strongest evidence because it's adversarial: arXiv 2607.28576 ran seven reflection-style methods against a plain repeated-sampling baseline on 1.5B/3B/7B open models, two math benchmarks, 150 questions each, counting every token spent on critiques, debate turns, and checking. No method reliably beat repeated sampling at matched cost anywhere. Ten were reliably worse. All 18 self-inspection comparisons came out negative. Reflexion as published never triggered its own retry on the smallest model: it judged itself correct every time and quietly collapsed into a single chain of thought.

That last detail should bother you. A self-critique layer that always self-approves looks identical, from the outside, to one that works.

Caveat honestly: that paper covers 1.5B–7B open models, not frontier models. The finding doesn't automatically transfer up. But combined with Anthropic telling you the same thing about its own flagship, the burden of proof has moved. Any self-critique layer you've bolted onto an agent loop needs a token-matched control before you believe it does anything.

Go delete something today. Start with /doctor.