A GitHub repo cuts 20K characters from Claude Code's always-on prompt for Opus 4.7. The override removes CAPS-heavy directives and anti-laziness scaffolding that the newer model handles worse.
Key facts
- 20K characters cut per coding turn from always-on prompt.
- CC ships every model the same prompt-by-volume Opus 4.6 needed.
- 41 net-new prompts caught by skrabe fork vs Piebald's JSON.
- 6 modifications derived from Opus 4.7 prompting guide.
- Repo targets CC 2.1.141, re-validated against each upstream release.
Anthropic ships Claude Code with the same system prompt volume for Opus 4.7 as it did for Opus 4.6. That's the problem.
A community repo called Lobotomized Claude Code, published by GitHub user skrabe, surgically rewrites the load-bearing fragments. The result: every coding turn carries roughly 20K fewer characters of always-on prompt. [According to Lobotomized Claude Code]
Why Opus 4.7 breaks under the old prompt
The repo's methodology reveals a structural mismatch. Anthropic's default prompt was tuned for Opus 4.6, which needed anti-laziness scaffolding and explicit instruction hierarchies. Opus 4.7 follows instructions more literally and overcorrects on CAPS directives like STRICTLY PROHIBITED and MUST. [According to the repo's prompting guide]
"4.7 is more literal — fewer rules, less drift," the guide states. "No CAPS theater." The override replaces strict prohibitions with plain directives.
The cold-path bulk that doesn't matter
Not all 20K characters come from active prompt. The repo identifies cold-path reference content — API docs for Python, TypeScript, Go, Java, C#, PHP, Ruby, the model migration guide, and skill-based cron tasks — that never injects in normal coding sessions. The headline cut focuses on the daily-flow row, where the always-on payload shrinks from roughly 45K to 25K characters.
Community tooling fills a vendor gap
The skrabe fork's extractor catches ~41 net-new prompts versus Piebald's published JSON for Claude Code 2.1.141. It also ships mechanisms — inline blob overrides with a string kind, a system-reminders runtime surface, and per-MCP-server routing — that reach dynamic injections the named-prompt pipeline alone can't touch. [According to Lobotomized Claude Code]
Hacker News commenters questioned whether Anthropic would ban users or move prompt logic server-side. The repo's existence suggests the default prompt is both observable and modifiable, which Anthropic has not explicitly sanctioned.
Six changes that matter
The repo distills six modifications from the Opus 4.7 prompting guide:
- Less is more: fewer rules, less drift.
- No CAPS theater: plain directives over
STRICTLY PROHIBITED. - Parallel by default: independent tool calls go in one message.
- Grill-me plan mode: the model interviews the user before producing a plan.
- No always-on CTAs: removed the
/scheduleupsell. - Tighter destructive-action guards: commit/push/merge/PR require explicit confirmation.
The unique take
This isn't just a prompt tweak — it's evidence that Anthropic's default Claude Code prompt is overfitted to Opus 4.6 and imposes unnecessary cognitive load on Opus 4.7. The community is doing what Anthropic should have done before shipping 4.7 support: re-validating the prompt against the model's actual behavior, not just copy-pasting the old one.
A similar pattern played out with the permission-first CLAUDE.md kit published on May 14, which addressed agent overreach. Both repos highlight a growing gap between Anthropic's release cadence and the prompt engineering required to make each model work optimally.
What to watch
Watch for Anthropic's next Claude Code release (likely 2.1.142+). If Anthropic silently adopts prompt reductions similar to the community overrides, it confirms the prompt was overfitted to Opus 4.6. If not, expect more community forks and a potential terms-of-service response.
