my claude code setup, open source. hooks, agents, tips, and a plugin that mines your usage data.
claude plugin add anipotts/mine # install the mine pluginthen: copy safety-guard.sh to block dangerous commands. read a tip. done.
hundreds of sessions across dozens of projects. $200/mo max plan.
same usage would cost ~$12K on the API with caching, ~$95K without. no autonomous loops. no cron jobs. every session starts with me typing a prompt. how the cost math works →
claude plugin add anipotts/mineyou get mine -- session mining to sqlite. costs, search, error memory, pattern detection. all data stays local at ~/.claude/mine.db.
/mine today's sessions, cost, top tools
/mine search "websocket" full-text search across all conversations
/mine mistakes error patterns claude keeps repeating
/mine hotspots most-edited files across sessions
/mine loops repeated patterns across sessions
start with mine + the safety-guard hook. add more as you go. mine docs →
hooks are the difference between "claude does what i want" and "claude does whatever it feels like." CLAUDE.md gives guidance. hooks give enforcement. one is a suggestion, the other is a wall.
i run 11 hooks on every session. safety-guard blocks force pushes, rm -rf /, and curl | bash. no-squash blocks squash merges. context-save preserves state before compaction. my full hook philosophy →
multiple claude instances working simultaneously on the same codebase, each in its own git worktree. the coordinator assigns tasks, collects results, merges the best approach.
i use this for parallel research, trying risky changes safely, and comparing approaches side-by-side without touching my working tree. how i use agent teams →
this is why the $200/mo plan is the best deal in AI coding. claude code caches your system prompt, tools, and CLAUDE.md as a prefix. 91% of my input tokens hit the cache -- meaning i pay 10% of the input cost on 91% of my reads.
the key: keep your CLAUDE.md short and stable. every edit breaks the prefix cache. mine is 30 lines and changes maybe once a week. the full cost breakdown →
short, standalone techniques. each one is something you can use in your next session.
| tip | what you learn |
|---|---|
| prompt caching | get 97%+ cache hit rates, slash your bill |
| ultrathink | force extended thinking for complex problems |
| settings hierarchy | project vs global vs local settings |
| safety hooks | block force pushes and rm -rf in 5 minutes |
| session length | why shorter sessions are more efficient (with data) |
copy one, wire it up, done. each is a standalone bash script. full guide →
| hook | event | what it does |
|---|---|---|
| safety-guard | PreToolUse | blocks force push, rm -rf /, DROP TABLE, curl-pipe-sh |
| no-squash | PreToolUse | blocks squash merges |
| panopticon | PostToolUse | logs every tool call to sqlite |
| context-save | PreCompact | saves context before compression |
| notify | Notification | routes to macOS, Slack, ntfy |
6 more hooks
| hook | event | what it does |
|---|---|---|
| commit-nudge | PostToolUse | reminds you to commit after N edits |
| replay-capture | PostToolUse | captures file changes for VHS replays |
| version-stamp | SessionEnd | auto-updates "tested with" stamps |
| stale-branch | SessionStart | warns about gone tracking branches |
| md-lint-fix | PostToolUse | auto-fixes markdown lint on save |
| knowledge-builder | PostToolUse | builds a codebase knowledge graph |
copy to .claude/agents/ and invoke with /agent <name>. each teaches a different pattern. guide →
| agent | pattern | what it does |
|---|---|---|
| watch-tests | daemon | watches files, runs tests, proposes fixes |
| try-worktree | worktree | tries risky changes in isolated worktrees |
| arch-review | quick review | fast architecture smell-test |
| write-pr | git integration | PR descriptions from your diff |
| command | what it does |
|---|---|
/mine |
usage data -- costs, sessions, search, patterns |
/ship |
stage, commit, push, open PR in one command |
/improve |
propose CLAUDE.md updates from git history |
plus 3 example commands you can copy: /sweep, /quicktest, /replay.
| what | |
|---|---|
| cost reality | what claude code actually costs, the prompt caching math |
| mistakes i made | what burned me so you can skip it |
| my stack | the automation setup i actually run |
| session workflow | how i work day-to-day with claude code |
| worktrees | parallel exploration with the desktop app |
diplomatic, data-driven, no FUD. every claim cites a source.
vs cursor · vs codex · vs gemini · vs antigravity · pricing
- CLAUDE.md templates -- starter configs for TypeScript, Python, Rust, Next.js
- example agents -- 4 agents, each teaching a different pattern
- example commands -- 3 commands you can copy to any project
- handoff plugin -- PreCompact context preservation
- broadcast plugin -- async notifications on git events
this repo runs on its own patterns.
- 12 CI workflows -- docs audit, competitive intel, community digest, freshness check, stale cleanup, dependabot, releases, plugin smoke test, PR quality gate, validation, claude responder, upstream watcher
- 11 hooks running on every session
- <$1/month CI cost -- AI-powered workflows use haiku
- 0 manual maintenance -- everything that doesn't require taste is automated
MIT · built by anipotts