Sabbir Amin
[NB. This is an opinionated post. There are other valid methods out there. In the agentic world, we learn every day, through mistakes and iteration. Not by luck ]
Different teams here have adapted differently, but the core stays the same. Here's how one of our teams does it.
It starts with design. We moved from Figma Make to a pure Claude Code + VS Code hybrid ( We started with Figma Make, but the continuous Limit on credits forced us to choose the alternative). Designer actually designs (vibe codes) in VS Code now, driving Claude, managing Git, and shipping React apps as prototypes. Dev team build the CI/CD pipelines, and the designer builds and deploys them regularly, so the product team can review and approve before development even starts.
Next comes development. The approved React app loads into Cursor. A clean project doc is generated from the code, then everything (PRD, API specs, Google Docs, third-party integration docs) gets fed into Claude's project mode to build a detailed knowledge base.
Two files anchor every project:
-
project.md- Think of it as the blueprint. What's being built, domain knowledge, folder structure. -
tasks.md: the task manager. Phases, milestones, blockers. Think of it as a Kanban board for agents.
These two files are the source of truth. Every session starts by binding to them. Every completed or blocked task gets updated in them.
Skills matter too. We maintain backend.md and
frontend.md files per project. Framework choices, version numbers,
specific libraries, best practices. All customized to the project's needs. For
teams not on Claude Code, these files plug into Cursor or whatever IDE they prefer.
Repo structure. Monorepo-like folder structure for agents, but each project has its own Git. Everything loads into a single workspace, commands run from root, agents navigate to backend or frontend as needed. Changes across both reflect instantly.
Context and token management. Each task ideally gets one session. After the session, project.md and tasks.md get updated so the next agent can pick up cleanly. We're also experimenting with a global context archiver, a vault for decisions and changes that informs future sessions. That probably deserves its own post.
Model choices. 🥲 Still experimenting, honestly. Cursor's Composer handles clear, straightforward tasks well. API implementation, UI updates. For heavier lifting, we rely on Anthropic and Google models. Sonnet has been the go-to combo. We're also testing Qwen and GLM. Learning as we go.
How are you adapting to agentic software development? I Would love to hear what's working or not working for your team.