lobotomized-claude-code

System-prompt overrides for Claude Code, tuned for Claude Opus 4.7. Targeting CC 2.1.141 (re-validated against each upstream release).

CC ships every model the same prompt-by-volume Opus 4.6 needed. 4.7 follows instructions literally, gets jumpier on CAPS, and doesn't need the anti-laziness scaffolding. This repo cuts the bulk and rewrites the load-bearing fragments in a register the model actually behaves better under.

Use skrabe/tweakcc-fixed (not the published tweakcc-fixed npm package) to apply these overrides. The skrabe fork's extractor catches ~41 net-new prompts versus Piebald's published JSON for CC 2.1.141, and ships additional mechanisms — the string kind in inlineBlobOverrides, a system-reminders/ runtime override surface, and per-MCP-server routing — that reach dynamic injections the named-prompt pipeline alone can't touch. Several of the overrides here depend on those mechanisms.

~67% off the daily flow. ~33% off everything.

What you actually feel	Cut
Daily flow — the fragments that inject every interactive turn (harness, communication, doing-tasks, executing-actions, memory, always-loaded tool descriptions)	~67% smaller
Niche stuff — language-specific Anthropic API docs, model migration guide, calendar/cron skills, WSL settings, Windows tool descriptions	~32% smaller

The first row is what matters. Niche prompts only inject when you trigger that feature (most users never do), so the more modest cut on those barely shows up in real sessions. Every coding turn carries roughly 20K fewer characters of always-on prompt. Faster first response, more headroom before compaction, less drift from contradictory always-on rules.

Methodology

Numbers measured against pristine CC 2.1.141 prompts. Anthropic doesn't publish an offline tokenizer for Claude 3+/4 models — accurate token counts require client.messages.count_tokens() against an API key — so figures are reported in characters (text length, no tokenization assumption). Same approximation isn't an issue when comparing two strings: the percentage cut is the percentage cut.

	Files	Pristine	After lobotomy	Cut
Hot path (always or near-always injected)	~24	~30K chars	~10K chars	−67%
Cold path (conditional / niche)	~267	~889K chars	~608K chars	−32%
Total	291	~919K chars	~618K chars	−33%

The cold-path bulk is reference content — data-claude-api-reference-{python,typescript,go,java,csharp,php,ruby}, the model migration guide, data-managed-agents-* API docs, the skill-* cron tasks (catch-up, dream, morning-checkin, pre-meeting). None of it injects in normal coding sessions, which is why the headline focuses on the daily-flow row.

What this aims at

Six things, all from the Opus 4.7 prompting guide:

1. Less is more. 4.7 is more literal — fewer rules, less drift.
2. No CAPS theater. STRICTLY PROHIBITED / CRITICAL REQUIREMENT / MUST triggers overcorrection on 4.7. Plain directives behave better.
3. Parallel by default. The harness now explicitly tells the model: independent tool calls go in one message, multiple blocks. Reading 5 files = 5 parallel reads. Stop running things in series when they don't have to be.
4. Grill-me on plan mode. Plan mode Phase 1 is now Matt Pocock's grill-me pattern — the model interviews you one decision at a time, recommends an answer, walks the design tree before producing a plan. Underspecified plans are more expensive than the questions.
5. No always-on CTAs. Killed the /schedule upsell prompt. Things 4.7 follows literally shouldn't be marketing copy.
6. Tighter destructive-action guards. Git commit/push/merge/PR all require explicit confirmation in the current conversation. "Commit and push" splits into two confirmations. The model can't make your DB go away because it got stuck.

Inspirations: pi and Factory Droid both ship full coding-agent system prompts in roughly the size of a single CC tool description. They work fine.

How it works

Each .md in system-prompts/ is an override fragment with frontmatter naming the original prompt and the CC version it targets. skrabe/tweakcc-fixed reads these and patches the cli.js (or native binary) of your installed Claude Code in place.

Three override mechanisms cover different injection paths:

• Named-prompt overrides — system-prompt-*.md, agent-prompt-*.md, tool-description-*.md, system-reminder-*.md, skill-*.md, data-*.md. These target named pristine prompts extracted into prompts-X.Y.Z.json. Largest bucket (~290 files), covers the bulk of CC's system prompt.
• Inline-blob overrides — inline-*.md. Target template-literals / string-arrays / quoted strings inside cli.js that aren't named pristine prompts. Reaches dynamic content that the extractor can't capture as a named entry (memory wrapper paragraphs, inline tool-usage bullets, skill descriptions for built-in skills, the # System bullets array, etc.). Each file's frontmatter carries a regex anchor and a kind (array / template / string).
• System-reminder overrides — ~/.tweakcc/system-reminders/<id>.md (different directory, separate mechanism in skrabe/tweakcc-fixed). Target per-turn / event-driven <system-reminder> dynamic injections — claudeMd context wrapper, telemetry, MCP per-server instructions, hook wrappers, etc.

Load-bearing structure (variable substitutions like ${ASK_USER_QUESTION_TOOL_NAME}, conditional blocks for env flags) is preserved verbatim. Anywhere you see ${...}, that's CC injecting runtime values — leaving them alone is non-negotiable.

Install

git clone https://github.com/skrabe/lobotomized-claude-code ~/.tweakcc/lobotomized-claude-code

# named-prompt + inline-blob overrides
mv ~/.tweakcc/system-prompts ~/.tweakcc/system-prompts.bak 2>/dev/null
ln -sfn ~/.tweakcc/lobotomized-claude-code/system-prompts ~/.tweakcc/system-prompts

# system-reminder overrides (per-turn / event-driven dynamic injections)
mv ~/.tweakcc/system-reminders ~/.tweakcc/system-reminders.bak 2>/dev/null
ln -sfn ~/.tweakcc/lobotomized-claude-code/system-reminders ~/.tweakcc/system-reminders

git clone https://github.com/skrabe/tweakcc-fixed ~/dev/tweakcc-fixed
cd ~/dev/tweakcc-fixed && pnpm install && pnpm build
node ~/dev/tweakcc-fixed/dist/index.mjs --apply

Per-MCP-server overrides are user-specific

tweakcc-fixed ships a per-MCP-server instruction-override mechanism: at --apply time it scans ~/.claude.json's mcpServers and routes each server's instructions field through ~/.tweakcc/system-reminders/mcp-<server-name>.md. Empty body drops that server's block from context, custom body replaces it, {{server_instructions}} passes the server's pristine text through.

This repo doesn't ship mcp-<server>.md files — MCP server lists are user-specific. After installing, write your own mcp-<server-name>.md files. They live outside the symlink (the install above only symlinks the system-reminders directory; you can break the symlink and manage that directory directly if you want both shipped defaults and personal MCP overrides), or you can fork this repo and add them to your fork.

A starter template for any MCP server:

<!--
name: 'MCP server: <your-server-name>'
description: per-server instruction override
ccVersion: 2.1.141
placeholders:
  - server_instructions
-->
{{server_instructions}}

Edit the body to lobotomize that server's instructions block. Empty body = server's block dropped entirely.

Re-run node ~/dev/tweakcc-fixed/dist/index.mjs --apply after every Claude Code update.

skrabe/tweakcc-fixed is a direct fork of Piebald-AI/tweakcc carrying unmerged patches that make --apply work on current CC versions, plus skrabe-only extraction extensions (more inline-blob kinds, the new system-reminders mechanism, per-MCP-server runtime routing). The tweakcc-fixed package on npm is a different fork (BenIsLegit/tweakcc-fixed) — these overrides are tuned against skrabe/tweakcc-fixed HEAD, so install from source above rather than via npx.

License

MIT. Derivative overrides of Anthropic's Claude Code prompts. Use on your local install at your own risk — these change the model's behavior in non-trivial ways and aren't endorsed by Anthropic.