The VC Framework.

The stack

The loop

Prompts are starting points — click any one to copy it. Adapt them to your repo; none are magic incantations.

The single biggest shift is the Model Context Protocol. Instead of copying documentation into the chat, the agent connects to MCP servers — GitHub, a database, your docs, Slack — and calls them as live tools. Anthropic introduced it, and OpenAI Codex and Google Gemini CLI support it natively too. If a step in your workflow is "paste the current API docs," an MCP server replaces it.

• Live, not stale The agent pulls current data the moment it needs it, instead of working from a snapshot you pasted an hour ago.
• One protocol, every agent The same MCP server works with Claude Code, Codex and Gemini CLI. Write the integration once.
• Tools, not text GitHub issues, database rows and search results arrive as structured tool calls the agent can act on — not a wall of pasted text.

• Trust Git, not the bot When the agent breaks something, revert — don't argue it into a fix. A clean commit per task makes that a one-liner. The most important rule here.
• Small, atomic tasks One reviewable change at a time. Small tasks fail cheaply, and they map cleanly onto subagents and plan mode.
• Tests are the guardrail Enforceable, not just advisory: a Stop hook can block the agent from finishing until the test suite is green.
• Reset context deliberately One task, one chat — but you do it with a command like /clear, not a new terminal, and auto-compaction covers the long runs.