Not a developer tutorial. Not a prompt engineering guide. This is a real, opinionated setup built by an artist who uses Claude Code every day for filmmaking, creative tools, and vibe coding.
I'm Ismaël Joffroy Chandoutis, a filmmaker and artist based in Paris. I work across cinema, contemporary art, and new media. I'm not a developer. I use Claude Code as a creative and technical partner to build tools, automate workflows, and explore AI as an artistic method.
This repo documents the actual configuration, decisions, and patterns that emerged from daily usage since January 2026. Everything here is tested in production (my production, which is making films and art).
How the agent is configured, and why.
| File | What it covers |
|---|---|
| CLAUDE.md structure | The global instructions file: identity, workflows, 7 quality rules (anti-hallucination, scrap-and-redo, self-updating rules, confrontation mode...) |
| Settings explained | Every settings.json choice annotated: privacy (telemetry off), deny rules (minimal: only block the irreversible), statusline, agent teams, extended thinking |
Key decisions:
- Deny rules are minimal by design: only block irreversible commands. The
claude(interactive permissions) vsclauded(bypass) separation handles daily control. - Telemetry fully disabled (Trail of Bits approach).
- Extended thinking always on.
- Conversation history kept 365 days instead of 30.
Reusable analysis, decision frameworks, and automation scripts.
| File | What it covers |
|---|---|
| Max Plan vs API | When Claude Code Max ($200/mo) pays off vs API. Real usage data: Jan $12, Feb ~$137 projected. Breakeven analysis. |
| tmux Survival Guide | Never lose a session. Multi-machine setup, remote access, recovery after reboot, navigation cheatsheet. |
| Telegram Bridge | Control Claude Code from your phone via Telegram. Bidirectional: text, voice, images, files. Multi-window support. |
| Notifications | Get notified when tasks finish. Sound (Zelda!) + push notifications (Pushover/Telegram/native macOS). |
| Statusline | Two-line status bar: model, project, context usage progress bar (color-coded), cost, duration, cache %, lines changed. |
| Scripts Toolkit | All automation scripts: bootstrap, dashboard, heartbeat, cost calculator, memory search, sync, nosleep, resume sessions. |
| Memory System | Persistent memory across sessions. Daily logs, backlog tracking, Spotlight search. Claude Code remembers what you did yesterday. |
| Resume Sessions | Restore all tmux windows after a reboot with claude --resume. One script, priority-ordered, named windows. |
| Cross-Platform | How everything works on macOS, Linux, and Windows (WSL2). Platform abstraction layer, bootstrap, and WSL2 setup guide. |
| Agent Layout Monitoring | Monitor subagent activity in real-time without leaving the main session. Split-pane layout, tmux hooks, live tail. |
| Multi-Machine Setup | MacBook Air M3 + Mac Mini M4 + PC Windows on Tailscale. SSH/mosh workflow, git sync strategy, MacBook = source of truth. |
| Ghostty + cmux | GPU terminal + native tmux workflow. Replaced iTerm2 (Feb 2026). Auto-restore all windows on reboot, no AppleScript. |
| Agent Teams | Peer-to-peer multi-agent coordination. When to use vs subagents. Task list sharing, messaging protocol, shutdown flow. |
Decisions log. What we tried, what worked, what we dropped.
| File | What it covers |
|---|---|
| Genesis | Day zero of systematic documentation. Full audit of our setup against gmoney.eth's 25 tips. Resulted in 7 new CLAUDE.md rules, statusline, deny rules, tmux aliases, daily memory logs. |
Writing published from this practice.
| File | What it covers |
|---|---|
| On Agentic Engineering | Notes from a filmmaker who's been through multiple technological ruptures in cinema — and recognizes this one. |
Articles and threads that shaped the setup.
| File | What it covers |
|---|---|
| gmoney.eth - 25 lessons | Annotated breakdown of @gmoneyNFT's thread (Feb 2026). Each tip scored: already implemented, implemented after reading, to explore later, decided against. |
| Karpathy — Vibe Coding | The original tweet that coined the term (Feb 2025, 6.7M views). Full transcript + why this repo is the attempt to make vibe coding sustainable rather than throwaway. |
AI as alteration, not augmentation. The agent isn't here to make me faster. It's here to make the work different from what I would have done alone.
Specs before code. Every project starts with an interview phase (20-30 questions for dev, 15-25 for creative). No implementation without specs. Two custom skills enforce this: /interview and /pitch.
7 quality rules baked into CLAUDE.md:
- Anti-hallucination -- never use simulated, invented, or approximate data
- Return to plan mode -- if a fix fails, stop. Don't spiral. Re-plan.
- Scrap and redo -- when output is mediocre, restart from scratch with accumulated context
- Self-update -- the agent updates its own rules after every significant error
- Use subagents -- parallelize complex tasks, keep main context clean
- Verify your work -- never say "done" without proof (tests, browser check, re-read)
- Confrontation mode -- challenge choices, say no when justified, don't just execute
- Max plan ($200/mo) — Opus 4.6 + Sonnet 4.6 + Haiku, unlimited usage (20x)
- 40+ official plugins active: code-review, frontend-design, superpowers, feature-dev, pr-review-toolkit, playwright, hookify, figma, Notion, Vercel, Sentry, Stripe, Pinecone, HuggingFace, Greptile, and more
- 109 skills — interview, pitch, video, SEO (6 dedicated agents), art generation, deployment, data analysis...
- 16 custom slash commands —
/interview,/pitch,/backlog,/briefing,/decision,/failure,/sync... - 38 persona agents — 32 filmmakers and thinkers (Tarkovsky, Akerman, Lynch, Duras, Foucault, Haraway...) + 6 SEO specialists. Each is a distinct creative voice for specific phases of work.
- 3 custom MCP servers: iMessage (read/search/send), Telegram (MTProto bridge), Postiz (social media scheduling API)
- Agent Teams (experimental) — peer-to-peer multi-agent coordination. Agents share a task list, assign work, communicate directly. Enabled via
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
- Ghostty — GPU-accelerated terminal, replaced iTerm2 (Feb 2026)
- tmux — persistent sessions, never lose work. Aliases:
za/zb/zc/zdfor window navigation - cmux — native tmux workflow replacing custom AppleScript layouts. Auto-restores all windows after reboot.
- MacBook Air M3 — primary workstation (Paris + travel)
- Mac Mini M4 — headless server, always on. Claude Code running 24/7 via tmux. Accessible via SSH + Tailscale from anywhere.
- PC Windows RTX 5090 — ComfyUI node for GPU-intensive generation (image, video, depth maps)
- iPhone — Telegram bot for remote Claude Code control (text, voice, images)
- iPad — Blink Shell (SSH to Mac Mini), Jump Desktop (remote desktop), full mobile workflow
- Tailscale — mesh VPN connecting all machines. MacBook, Mac Mini, PC Windows, iPhone, iPad. SSH to Mac Mini from anywhere:
ssh ismaelstudiominim4@100.84.223.88 - Mosh — stable SSH over mobile networks (Tailscale + mosh combo)
- Postiz on Mac Mini — social media scheduling, self-hosted via OrbStack/Docker (7 containers). API at
http://100.84.223.88:4007. - ComfyUI on PC — local inference for Wan 2.1 T2V/I2V, SDXL, RIFE interpolation, ESRGAN upscale. Accessible via Tailscale.
- Comfy Cloud — overflow GPU ($20/mo, 96GB VRAM) for heavy workflows.
- Hooks: Zelda sound + Telegram push on every task completion
- Heartbeat agent — periodic background checks, health monitoring
- Auto-sync cron —
~/.claude/synced to GitHub every 30min (MacBook) / 1h (Mini) - VoiceInk — offline voice dictation
- BACKLOG.md — single source of truth for all open work across sessions
memory/YYYY-MM-DD-subject.md— session logs: decisions, problems solved, todos- Spotlight search —
mdfind -onlyin ~/.claude/memory/ "keyword"for instant recall
- Trail of Bits -- opinionated security defaults
- @bcherny (Claude Code creator) -- team tips
- @gmoneyNFT -- 25 lessons from daily multi-agent usage
- @__BOMO -- context window statusline
Living document. Updated after every significant session or setup change.
This is not a template to copy. It's a reference for people building their own relationship with an AI coding tool, especially non-developers using it for creative work.