Explore, plan, code, commit
A four-step workflow for coding agents: read the relevant code, write a plan, implement it, then commit with a clear message.
- Level
- Advanced
- Time
- About 30 min
- Works with
- Claude Code, Other coding agents
Why it is here
It is the simplest change that improves AI coding results. The agent stops guessing and starts reading.
How to use it
- 1 Explore: ask the agent to read the relevant files and not write code yet.
- 2 Plan: ask for a short plan and review it yourself.
- 3 Code: let it implement the plan and run the tests.
- 4 Commit: ask for a commit and, if needed, a pull request with a summary.
The pattern behind it
Use the same idea for your own tasks.
Did this work for you?
- #claude-code
- #agents
- #best-practice
Related
- →
Anthropic Agent Skills
Primož's pick NewThe official collection of Agent Skills from Anthropic: folders of instructions and scripts that Claude loads only when a task needs them (documents, spreadsheets, slides, skill creation).
🧩 Skill 💻 Coding Advanced ~20 min · GitHub · anthropics - →
Superpowers
NewA community skills library for coding agents with skills for brainstorming, planning, test-driven development and systematic debugging.
🧩 Skill 💻 Coding Advanced ~15 min · GitHub · obra - →
MCP reference servers
NewReference servers for the Model Context Protocol (filesystem, git, fetch, memory and more). They show how an AI assistant connects to tools and data.
🧩 Skill 💻 Coding Advanced ~30 min · GitHub · modelcontextprotocol