The foundational Claude Code workflow from Anthropic's official best practices guide. Every community workflow listed in DEVELOPMENT WORKFLOWS is a specialization of this four-phase pattern.
| ← Back to Claude Code Best Practice |
| Phase | Goal | What Claude does | How to enforce |
|---|---|---|---|
| Explore | Build context. No edits. | Reads files, runs grep/find, gathers facts. |
Say "do not write code yet" or delegate to the Explore subagent. |
| Plan | Design before doing. | Proposes an approach you can critique. | Plan Mode (Shift+Tab twice), or /ultraplan. |
| Code | Execute the approved plan. | Edits files according to the plan. | Switch out of Plan Mode only after the plan is reviewed. |
| Commit | Capture the change cleanly. | Creates focused, reviewable commits. | One commit per file (see project rule). |
The failure mode it prevents: Claude reads two files, infers a pattern, writes 200 lines, then discovers it misunderstood — now you have a half-correct diff to untangle.
Splitting Explore and Plan from Code makes that misunderstanding visible before code exists. The user reviews a plan (cheap to revise) instead of a diff (expensive to revise).
Two of the global behavioral guidelines in ~/.claude/CLAUDE.md map directly onto the first two phases:
- §1 Think Before Coding = Explore + Plan
- §4 Goal-Driven Execution = Plan defines verifiable success criteria before Code starts
Each phase has tooling built into this repo's configuration:
| Phase | Built-in support |
|---|---|
| Explore | Explore subagent (parallelizable — spawn multiple in one message for independent searches) |
| Plan | Plan Mode, /ultraplan, EnterPlanMode tool |
| Code | Default mode; Auto Mode for less prompting |
| Commit | Per-file commit rule, Checkpointing for safe rewinds |
- TDD variant: Explore → Plan → Test → Code → Commit. Write a failing test after Plan; Code runs until tests pass. Common in the Superpowers and Matt Pocock workflows.
- Visual iteration variant: For UI work, add screenshots after Code. Use the Playwright MCP or
agent-browserskill. - Cross-model review variant: Add a second-model review between Plan and Code (see cross-model workflow).
- Parallel Explore: Spawn multiple
Exploresubagents in one message for independent searches — recommended in the Agent tool docs.
| Anti-pattern | What goes wrong |
|---|---|
| Skip Explore — start in Code | Claude infers structure from filenames, gets it wrong, edits the wrong file. |
| Skip Plan — bundle Explore + Code | You review a 200-line diff instead of a 5-line plan. Expensive to revise. |
| One giant commit at the end | Lost reviewability. Hard to bisect. Violates this repo's per-file rule. |
| Re-Explore mid-Code | Context bloats. Use /compact with a hint or /rewind instead. |
- Claude Code Best Practices — original Anthropic post
- Common Workflows — official docs
- DEVELOPMENT WORKFLOWS — community workflows that extend this pattern
tips/claude-boris-13-tips-03-jan-26.md— Boris Cherny on plan mode discipline