Claude Code Config

Opinionated Claude Code defaults for cost-aware, model-tiered development on Amazon Bedrock.

Default Claude Code sessions inherit the main-loop model for every subagent, silently multiplying cost on frontier sessions. This config:

• Pins Sonnet 4.6 as the everyday default (cheap, capable, adaptive thinking)
• Makes Opus 5, Fable 5, Haiku 4.5, and an opusplan hybrid available one /model keystroke away — escalating to frontier reasoning is deliberate, never accidental
• Tiers subagents by value in CLAUDE.md so mechanical tasks run on Haiku and only judge/synthesis work runs on Opus
• Pins exact Bedrock inference-profile IDs (on Bedrock, bare aliases like opus resolve to older model versions without these pins)

git clone https://github.com/kaltura/claude-code-config.git
cd claude-code-config

cp CLAUDE.md ~/.claude/CLAUDE.md
cp settings.json ~/.claude/settings.json
cp managed-settings.opusplan.json \
  "/Library/Application Support/ClaudeCode/managed-settings.json"

cp -r git-hooks ~/.claude/git-hooks
git config --global core.hooksPath ~/.claude/git-hooks

Not on Bedrock? Remove the ANTHROPIC_DEFAULT_*_MODEL env vars from managed-settings.opusplan.json. Bare aliases (sonnet, opus, etc.) will resolve against the Anthropic API directly.

Telemetry: settings.json includes OTLP env vars pointed at http://127.0.0.1:10198. Remove or update if you don't have a local collector. The OTEL_EXPORTER_OTLP_HEADERS field is where an auth token would go — never commit a value there.

Plugins: the enabledPlugins and extraKnownMarketplaces entries in settings.json are personal. Remove them or replace with your own. The ones registered here:

Attribution: settings.json sets attribution.commit and attribution.pr to hide the noreply@anthropic.com email, keeping just Co-Authored-By: Claude Code. A known Claude Code bug (anthropics/claude-code#45137) can make the commit-trailer setting get ignored, so git-hooks/commit-msg is a backup that rewrites the trailer after the fact. core.hooksPath is a global git setting: only install it if you don't already rely on per-repo hooks, since it replaces core.hooksPath for every repo on the machine that doesn't set its own. There's no equivalent backup for PR descriptions, since gh pr create doesn't go through git hooks.

claudeMdExcludes: settings.json excludes /opt/homebrew/CLAUDE.md from auto-loading. This is a personal filesystem-layout fix, not a general recommendation: on Apple Silicon, brew --prefix is /opt/homebrew, and Homebrew commits its own contributor CLAUDE.md/AGENTS.md at that repo's root. Claude Code walks up the directory tree and loads every ancestor CLAUDE.md it finds, so any project nested under /opt/homebrew (e.g. /opt/homebrew/var/www/...) picks up Homebrew's Ruby/Sorbet/RuboCop guidance by accident. Excluding the one file also suppresses AGENTS.md, since it's only pulled in via an @AGENTS.md import inside that CLAUDE.md — excluding the importer means the import never fires. If your projects don't live under /opt/homebrew, remove this key entirely.

The data and reasoning behind the defaults in this repo:

• ai-cost-efficiency-learnings.md — how one deployment cut daily spend 86% through model tiering and cache discipline
• model-tiers-vs-harness-guide.md — when to fix your prompt/harness, when to raise effort, and when to actually escalate model tier
• lessons-from-re-auditing-my-own-cost-report.md — four lessons from auditing whether those savings held up over time, including a caching pitfall that quietly erased a cheaper model's rate-card advantage