Olivier Elemento, Director of Englander Institute for Precision Medicine at Weill Cornell Medicine, shared a post on LinkedIn:
“My AI stack for coding, research and writing.
There’s been great discussion lately about AI stacks – Peter Steinberger on shipping at inference speed, Ben Tossell on coding with agents without being ‘technical’.
My use case is different: research, writing, grant management, email, talks.
And coding of course – something I more or less stopped doing for many years, but that now Claude Code makes feasible again (and fun).
Here’s what I use daily:
Voice input via Wispr Flow. I talk to my computer more than I type.
Voice captures thinking faster than typing, and modern speech recognition handles technical terms and accents well.
I still use the keyboard when I can’t speak out loud, but I think voice will dramatically change how we interact with machines.
Multiple concurrent AI agents. I run one Terminal window with multiple tabs, each running Claude Code, Gemini, or Codex.
I call each one an ‘agent’ because it has persistent context for a specific role: Personal Assistant, Project Manager, Culture Coach, LinkedIn posts, or a coding project. I wrote my own manager to orchestrate them – fast tab switching, status indicators showing whether a model is working or waiting.
Most agents run Claude Code with Opus 4.5 (I am a Max subscriber), plus 1-2 Gemini agents when long context is needed.
The agents can draft emails and open them in Outlook ready to send, add calendar items and send text messages via AppleScript, generate figures by calling the Gemini Nano Banana Pro API (for example, high-quality Gantt charts).
Claude skills and plugins for task context.
These are reusable instruction sets – how to read/write word documents (including putting documents on letterhead), draft LinkedIn posts, draft emails, draft paper reviews.
The agent loads the relevant skill and follows it.
For new coding projects, the Superpowers Brainstorming skill has been particularly helpful – it forces exploration before implementation.
I am gradually splitting my long CLAUDE.md instruction file into skills: Markdown files loaded on demand rather than always present.
Obsidian vault as the knowledge base. All my daily notes, projects, contacts, document drafts live there. The agents read from it and write to it.
I use git to sync it to my cloud server (VPS), and iCloud to sync to my phone. Obsidian is just Markdown files – simple, natively AI-compatible, and far better than Evernote ever was.
Cloud server with Termius for remote access. This lets me run one Claude Code instance attached to my Obsidian notes from anywhere.
The VPS updates the vault via git, so changes sync back to my laptop.
The insight I keep coming back to: the stack is less about any single tool and more about the connective tissue.
Voice captures ideas, agents process them, the vault stores context, skills ensure quality. Each piece amplifies the others.”
More posts featuring Olivier Elemento.