rawcode

rawcode

OpenCode's philosophy for Claude Code.
Install it, and Claude Code thinks different.

Install · What Changes · Guardrails · Changelog

---

No commands. No agents to pick. Just a system prompt that makes Claude code better.

# Install in one line
curl -fsSL https://raw.githubusercontent.com/juandarn/rawcode/master/setup/install.sh | bash

What Changes

rawcode injects a system prompt into every Claude Code session. You don't invoke anything — just open claude and it works.

Principle	What Claude does	Without rawcode
Read first	Reads existing code before writing	Generates blind, duplicates utilities
Concise	Responds in < 4 lines	20 lines of explanation, 5 of code
Root cause	Fixes the origin, not the symptom	Patches symptoms that come back
Minimal	Only changes what's necessary	Drive-by refactoring, extra features
Secure	OWASP checklist on every change	"Be cautious with user input"
Verify	Runs lint and tests after every change	Says "done" without checking
No fluff	No preamble, no summaries, no praise	"Great question! Let me explain..."

Everything stays native — /plan, /compact, and all Claude Code features work as usual.

How it compares

Measured, not claimed. rawcode's prompt was A/B-tested against the Claude Code baseline on HumanEval — a standard dataset graded by executable unit tests (no LLM judge, no bias). Paired design, claude -p, single greedy sample per problem, n=40. Correctness and length are reported as separate axes (a length-biased judge would conflate them):

Axis	Baseline	rawcode	Paired Δ (95% CI)
Correctness — pass@1	90.0%	95.0%	+5.0% [0, +12.5%] · McNemar p=0.50 → no significant change
Output tokens / problem	229	150	−35% [−138, −34] → significant

Honest reading: on this code-generation set, rawcode cut output by about a third with no measurable change in correctness — terseness was free here. It does not claim higher quality (the correctness difference is within noise), and it does not test long agentic loops, where Anthropic has reported that over-aggressive between-step conciseness can hurt quality — which is exactly why rawcode's brevity rule binds to the final response, never to investigation or verification.

Caveats: HumanEval is partly present in training data (a paired A/B cancels most of this since both arms see the same problems); vanilla tests are weaker than HumanEval+; n=40 single-sample means the correctness CI is wide while the token effect is robust. Reproduce it yourself: bench/.

Install

Marketplace (recommended)

claude plugin marketplace add juandarn/rawcode
claude plugin install rawcode@rawcode

Guardrails and the TDD gate load automatically. To turn on the terse coding persona: /output-style rawcode.

Mac / Linux (script)

curl -fsSL https://raw.githubusercontent.com/juandarn/rawcode/master/setup/install.sh | bash

Windows

irm https://raw.githubusercontent.com/juandarn/rawcode/master/setup/install.ps1 | iex

Manual

git clone https://github.com/juandarn/rawcode.git ~/.claude/plugins/rawcode

Framework

rawcode is minimal by default, but it grows into a coding framework when you want one — a strict TDD gate, a rc-tdd red→green→refactor skill, and a real bench/ pass@1 harness. It works standalone or as an extension of Gentle AI, composing with its SDD flow instead of duplicating it. See docs/FRAMEWORK.md for the architecture and docs/COMPOSE.md for using the two together.

Guardrails

Automatic protections that run without you doing anything:

Guardrail	What it does
Protect sensitive files	Blocks editing .env, migrations, lock files
Read before write	Reminds Claude to read files before modifying them
Sanitize commits	Strips auto-generated attribution from commit messages

How it Works

rawcode is a Claude Code plugin. It has four parts: