---

Claude Code feeds every model the same prompts — written for older Claudes and padded with scaffolding, hedging, ALL-CAPS warnings, and anti-laziness nagging that newer models don't need and often behave worse under. This repo catalogues every one of them, rewrites the load-bearing ones in a register the model behaves better in, and deletes the parts that earn nothing. skrabe/tweakcc-fixed splices the result straight into your installed Claude Code.

The cut

Live Opus 4.8 set, measured against Claude Code 2.1.187:

Surface	Stock	Lean
System prompts	159K	99K	−38%
Tool descriptions	153K	105K	−31%
Everything	1.74M	1.26M	−28%

That's ≈488,000 characters gone — 435 prompts rewritten tighter and 66 cut outright (unused features, always-on upsells, duplicated warnings). The system prompt and tool descriptions are what the model re-reads on every request, so a third off them is a third less noise each turn: a quicker first token, more headroom before compaction, and fewer contradictory rules pulling against each other.

The one set that's live

system-prompts-opus-4-8/ — reshaped from scratch against the Opus 4.8 system card, and kept in lockstep with tweakcc-fixed's extractor. If you're on Opus 4.8, this is the set.

The rest is a legacy shelf: system-prompts-opus-4-7/ for Opus 4.7 / 4.6, and system-prompts-fable-5/, frozen now that the Fable line is retired. 🫗

What "lobotomized" means

One rule, run over every prompt: cut what doesn't earn its tokens, keep what changes behavior, leave the user-facing copy alone. It's a judgment pass against the model's system card, not a word-count diet — each prompt is re-read and rewritten, or dropped, on its own merits.

• Cut — anti-laziness scaffolding, CAPS theater (MUST / NEVER / ALWAYS), reflexive hedging and caveats, always-on upsells, restated rules, and docs for features you don't use (Managed Agents, PowerShell, WSL). Things 4.8 either does by default or behaves worse under.
• Keep, sharpened — the honesty contract (report failures straight, never fake success), destructive-action confirmation, scope discipline, parallel-tool guidance, exact output formats.
• Strengthen — the one surface the system card flags 4.8 as weaker on: browser / computer-use / web-fetch. Treat page and tool output as untrusted, scrutinize intent, confirm before anything destructive.

Install

These are overrides, not a tool — tweakcc-fixed applies them.

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

# point tweakcc at the Opus 4.8 set, plus the per-turn reminders
ln -sfn ~/.tweakcc/lobotomized-claude-code/system-prompts-opus-4-8 ~/.tweakcc/system-prompts
ln -sfn ~/.tweakcc/lobotomized-claude-code/system-reminders        ~/.tweakcc/system-reminders

# patch your installed Claude Code
npx -y tweakcc-fixed@latest --apply

Re-run --apply after each Claude Code update; --restore puts the originals back. Switching sets later is one ln -sfn.

How it works

Each .md is one prompt. An HTML-comment header carries the metadata — its id, the Claude Code version it was cut against, the ${VARIABLES} it interpolates — and everything below is the replacement text.

<!--
name: 'System Prompt: Doing tasks'
ccVersion: 2.1.187
-->
When the user asks you to do something, just do it. ${TASK_GUIDANCE}

An empty body suppresses the prompt outright — the model never sees it. A ${VAR} placeholder leaves Claude Code's runtime interpolation intact. tweakcc-fixed matches each file to its pristine prompt by id and splices the new text in. The per-turn <system-reminder> injections — the ones that never surface as named prompts — get the same treatment in system-reminders/.

Credit & license

Derivative overrides of Claude Code's prompts (© Anthropic), applied with skrabe/tweakcc-fixed. They change model behavior in non-trivial ways and aren't endorsed by Anthropic — run them on your own local install at your own risk. MIT.