My dev and VC research workflow has shifted quite a bit over the past month. I barely use Cursor agent anymore - just autocomplete. Most tasks are handled by Claude Code, usually with 2 or more agents running in parallel. Here’s a typical project flow: - Gemini 2.5 Deep Research for digging into the problem background, analyzing 100+ pages and articles on the topic, and writing a full report - Grok-4 for deep conceptual work - mapping out the problem space, writing the PRD or a requirements doc with expected outcomes - Taskmaster (screenshot) for project planning - breaking a big project into 20–50 tasks - Creating and /docs - the most important step in my view: holds the core philosophy and working principles of the project (for code, every rule and edge case; for research, all constraints and methods); /docs holds detailed AI-generated instructions for every major step - If the task is easy to parallelize, I use something like claude-flow to spin up a swarm or hivemind of dozens of agents working simultaneously - After every significant milestone - a GitHub commit, like a save point in a video game - Claude tokens usually run out (my $100/month plan isn’t enough anymore), but you can plug Claude Code into Kimi via Groq