How I use Claude Code as a filmmaker and contemporary artist working at the intersection of cinema, art, and AI.
This is a living document of my Claude Code configuration. Not a tutorial, not a template. A working setup that evolves daily.
| Component | Count |
|---|---|
| Official plugins | 40+ (Notion, Vercel, Sentry, Figma, Playwright, HuggingFace...) |
| Custom skills | 109 |
| Agent personas | 38 (filmmakers, philosophers, artists) |
| Custom slash commands | 16 |
| Custom MCP servers | 3 (iMessage, Telegram, Postiz) |
| Custom scripts | 30+ (notifications, statusline, sync, heartbeat, backup) |
| Inter-session memory files | Growing daily |
~/.claude/
├── CLAUDE.md # Master instructions (the brain)
├── BACKLOG.md # Centralized project tracker
├── settings.json # Hooks, statusline, permissions
├── agents/ # 38 creative consultant personas
├── skills/ # 109 skills (custom + plugin-managed)
├── scripts/ # Automation (notifications, sync, etc.)
├── memory/ # Inter-session persistence
├── sounds/ # Notification sounds
└── secrets/ # API keys (gitignored)
Claude Code is not an assistant. It's a collaborator that alters rather than augments. It anticipates needs, produces proactively, analyzes patterns. No productivity coaching, no hustle, no sycophancy.
Key rules enforced in my config:
- Anti-hallucination: Never use simulated, invented, or approximate data
- Scrap and redo: When output is mediocre, restart from scratch with accumulated context rather than patching
- Confrontation mode: Challenge my choices when relevant. I'm not a dev, I need critical thinking, not passive execution
- Self-updating rules: After every significant error, the config updates itself to prevent repeats
- One session = one subject: No mixing accounting, SEO, and creative work in the same session
38 creative consultant personas spanning cinema, philosophy, art theory, and game design. Each persona has a distinct voice, methodology, and area of expertise.
They can be invoked individually for creative consultation or debated against each other via a /council command (in development).
The personas are private by design. The concept is public: a system that lets you think through creative decisions with the worldviews of artists and thinkers you admire.
Mix of custom-built and plugin-managed skills. Organized by domain:
cw-brainstorming cw-prose-writing cw-story-critique cw-style-skill-creator cw-official-docs cw-router
pitch (creative specs: films, games, installations, performances) brainstorming (mandatory before any creative work) transcribe speech pdf doc
fal-generate fal-image-edit fal-upscale fal-audio fal-workflow fal-platform imagegen sora nanobanana logo-creator banner-creator canvas-design image-manipulation-imagemagick video-downloader
3d-web-experience develop-web-game game-development motion-canvas remotion remotion-best-practices algorithmic-art
frontend-design interactive-portfolio scroll-experience claude-d3js-skill excalidraw-diagram-generator web-artifacts-builder web-design-reviewer webapp-testing
seo seo-audit seo-page seo-geo seo-competitor-pages seo-hreflang seo-images seo-sitemap seo-plan seo-schema seo-technical seo-content seo-programmatic
data-query reddit twitter x-research producthunt requesthunt domain-hunter agentic-eval
dispatching-parallel-agents executing-plans finishing-a-development-branch gh-address-comments gh-fix-ci git-commit jupyter-notebook mermaid-diagrams playwright receiving-code-review requesting-code-review refactor skill-creator subagent-driven-development systematic-debugging verification-before-completion writing-plans
security-best-practices security-ownership-map security-threat-model
briefing inbox-sort linear notion-knowledge-capture notion-meeting-intelligence notion-research-documentation notion-spec-to-implementation prd
cloudflare-deploy netlify-deploy render-deploy vercel-deploy yeet
pdf pdftk-server pptx spreadsheet screenshot mcp-builder
status backup sync bootstrap cleanup skill-creator mcp-cli compta
I don't code directly. Every project follows this workflow:
1. /pitch [type] → 15-25 creative questions
(film, game, installation, performance, workshop)
2. Review & mature → Adjust, let it sit
3. /interview [specs.md] → 20-30 technical questions
4. Review specs → Adjust implementation details
5. Plan Mode → Claude proposes architecture
6. Implementation → Code with full context
If I try to jump straight to code, Claude stops me and redirects to /interview.
- iMessage - Full history, search, contact enrichment, send
- WhatsApp - Chats, contacts, messages
- Discord - Servers, channels, forums, webhooks
- Telegram - Full account access (MTProto), group management, leave/join
- Google Chat - Spaces, threads, DMs
- Gmail - Search, read, send, labels, drafts
- Google Calendar - Events, free time, scheduling
- Google Drive - File search, download
- Google Sheets/Slides/Docs - Read, write, search
- Notion - Pages, databases, blocks, comments
- Apple Notes - Read, search
- Framer - Direct site editing via MCP (project XML, node manipulation, CMS, code files, styles)
- Chrome DevTools - Screenshots, console, network, performance tracing
- Obsidian - Vault access, search, cache
- Brave Search - Web, news, images, video, local
- Crypto Market - Real-time data (Crypto.com Exchange)
Task completion triggers:
- Zelda treasure chest sound (local
afplay) - Push notification via Pushover (reaches iPhone/iPad instantly)
# ~/.claude/scripts/notify.sh
afplay ~/.claude/sounds/treasure.wav &
curl -s -F "token=${PUSHOVER_TOKEN}" -F "user=${PUSHOVER_USER}" \
-F "message=${MESSAGE}" -F "title=Claude Code" \
https://api.pushover.net/1/messages.jsonBoth Stop and Notification hooks are wired to this script.
A custom two-line status bar showing real-time session info:
[Opus] project-name │ main
████████░░░░ 67% │ $2.41 │ 34m ↻89% │ +124-38
Line 1: Model, project folder, git branch Line 2: Context usage (color-coded progress bar), session cost, duration, cache hit rate, lines changed
The full script parses Claude Code's JSON status input with jq and formats it with ANSI colors.
Full Claude Code access from mobile via tmux + Tailscale + mosh:
Mac (always-on) iPhone/iPad
┌─────────────┐ ┌──────────────┐
│ tmux session │◄────────►│ mosh client │
│ └─ claude │ mosh │ (Blink/a-Shell)
│ │ over │ │
│ Tailscale │ Tail- │ Tailscale │
│ 100.x.x.x │ scale │ 100.x.x.x │
└─────────────┘ └──────────────┘
OAuth stays active as long as the tmux session lives. Detach/reattach seamlessly.
Every significant session saves a summary to ~/.claude/memory/YYYY-MM-DD-subject.md:
- Decisions made
- Problems solved (and how)
- Things learned
- TODOs for next session
~/.claude/BACKLOG.md is the single source of truth for all active projects. Consulted at session start, updated at session end. Currently tracking 22 active workstreams and 14 completed.
mdfind -onlyin ~/.claude/memory/ "keyword"This isn't a developer's config. It's built for someone who:
- Makes films and contemporary art, not SaaS products
- Needs creative consultation, not code review
- Works across cinema, installation art, digital art, and games
- Treats AI as a collaborator in the artistic process, not a tool for productivity
- Manages complex multi-year film productions alongside daily comms and admin
The 38 agent personas aren't a gimmick. When I'm working on a documentary and need to think about the ethics of re-enactment, or when I'm designing an installation and need to consider the relationship between body and screen, having specialized creative perspectives available in the workflow changes the quality of the thinking.
- MacBook Air M3 (primary)
- Mac Mini (sync target)
- PC Windows (ComfyUI / GPU workloads via Tailscale)
- MacBook Air M3 — primary workstation (Ghostty + tmux, replaced iTerm2 Feb 2026)
- Mac Mini M4 — headless server, always on (Paris), Claude Code 24/7
- PC Windows RTX 5090 — ComfyUI GPU inference, accessible via Tailscale
- iPhone/iPad — Blink Shell + mosh for mobile Claude Code access
- Ghostty — GPU terminal, replaced iTerm2 (Feb 2026)
- Postiz — self-hosted social scheduling on Mac Mini (OrbStack/Docker)
- Tailscale — mesh VPN connecting all machines
- Agent Teams — peer-to-peer multi-agent coordination (experimental)
- 40+ official plugins active
- 109 custom skills
- 38 agent personas
- 3 custom MCP servers + extensive plugin MCP ecosystem
- 5 messaging platforms accessible
- Multi-machine setup (MacBook + Mac Mini + PC Windows)
- Custom hooks, statusline, memory, remote access
- All running from the terminal, no IDE needed
Ismaël Joffroy Chandoutis Filmmaker & Artist | Paris ismaeljoffroychandoutis.com