darkroom

Orchestrate multi-agent AI teams to build features, refactor codebases, audit security/performance, generate tests, conduct code reviews, and ship GitHub PRs for Next.js/React apps using TypeScript, Tailwind, Biome, and Vitest. Manages persistent memory, context handoffs, and standards enforcement for complex, long-running workflows.

Darkroom Claude Code Settings

Engineering-grade configuration for Claude Code, Codex, and other AI coding agents. Not vibe coding.

A composable layer of standards, agents, skills, and hooks that prevents the failure modes that show up when you actually try to ship with AI. Used in production at darkroom.engineering.

Open source. Hack around. Make it your own.

Quickstart (30-second setup)

bash <(curl -fsSL https://raw.githubusercontent.com/darkroomengineering/cc-settings/main/setup.sh)

Restart Claude Code. That's it.

Re-runs are non-destructive — your hand-added permissions, custom hooks, and local env overrides all survive. See docs/settings-reference.md.

Alternative install methods

Clone and run locally:

git clone https://github.com/darkroomengineering/cc-settings.git /tmp/darkroom-claude
bash /tmp/darkroom-claude/setup.sh

Interactive install (prompts on each settings conflict):

bash setup.sh --interactive    # or: CC_INTERACTIVE=1 bash setup.sh

Plugin install (requires GitHub SSH keys):

/plugin install darkroomengineering/cc-settings

Windows:

.\setup.ps1

Why cc-settings exists

These are the four failure modes we kept hitting on real projects, and the fixes that ship with cc-settings.

#1: Agents drift from the codebase

"With a ubiquitous language, conversations among developers and expressions of the code are all derived from the same domain model."

Eric Evans, Domain-Driven Design

The problem. Default Claude Code (and Codex, Cursor, Copilot) sees only the conversation. After a few turns, the agent forgets project conventions, hallucinates function names, or applies patterns from its training data instead of yours. Without a shared reference, every session starts from scratch.

The fix. A layered context system that loads contextually instead of all at once:

AGENTS.md — portable coding standards. Read by Claude Code, Codex, Cursor, and 20+ other tools (the open standard)
rules/ — path-conditioned guidance. react.md loads when editing .tsx, git.md always loads, security.md only on API/lib code
profiles/ — opt-in stacks for Next.js, Expo, Tauri, WebGL

Per-agent context cost dropped from ~5,000 tokens (one giant CLAUDE.md) to ~2,700 tokens by splitting and gating. See the table under Architecture.

#2: Sessions hallucinate fixes that don't compile

"Always take small, deliberate steps. The rate of feedback is your speed limit. Never take on a task that's too big."

David Thomas & Andrew Hunt, The Pragmatic Programmer

The problem. AI agents generate code that looks plausible but doesn't typecheck, doesn't run, or breaks tests three files away. They confidently report "fixed" when the build is red.

The fix. Multiple feedback loops, all enforced:

Pre-commit tsc hook — blocks any commit if types fail. No --no-verify shortcuts
/verify — three-agent adversarial pattern (finder → adversary → referee) for high-stakes code (auth, payments, data integrity)
2-iteration limit (in AGENTS.md) — fail twice, stop, present alternatives instead of digging deeper into a doomed approach
/fix — explore → tester → implementer → reviewer pipeline. Never solo
/ship — typecheck → build → test → lint → web quality gate before any PR

#3: Context window burns out before the work is done

"Premature compaction is the root of all unfinished features."

The problem. Non-trivial features can blow through the 1M-token Opus context before shipping. Auto-compaction at 95% loses state. Sessions die with WIP unsaved. Subagents reload the same config 9× across an orchestration.

The fix. Token-aware everything:

Layered config loading keeps base cost at ~2,700 tokens (was ~5,000). Subagents inherit, don't reload
Prompt caching 1h (ENABLE_PROMPT_CACHING_1H) — cache hits cost ~10% of fresh
Native Skill tool auto-invocation (Claude Code v2.1.108+) — skill text only loaded when the description matches your prompt
/checkpoint mid-session JSON quicksaves; /create-handoff for full session transfers
Auto-handoff on PreCompact — never lose work to compaction. Resume next session with /resume-handoff

#4: Upstream changes leave configs stale

"The only constant in agent tooling is the changelog."

The problem. Claude Code ships ~3 versions per week. Native features start overlapping with custom hooks. Schemas drift from reality. You either chase the changelog manually every Friday or fall behind and find out the hard way.

The fix.

darkroom

Component Overview

Component Details

Agents (10)

Skills (42)

Help us improve

Help us improve

darkroom

Component Overview

Component Details

Agents (10)

Skills (42)

Help us improve

README

Darkroom Claude Code Settings

Quickstart (30-second setup)

Why cc-settings exists

#1: Agents drift from the codebase

#2: Sessions hallucinate fixes that don't compile

#3: Context window burns out before the work is done

#4: Upstream changes leave configs stale

Similar Plugins

web-dev

everything-claude-code

cc-godmode

prompt-mini

claude-code-agents

oac