Claude Code — Workflows · AI coding agents

Tool-specific mini-workflows built from verified commands. For cross-tool versions of these workflows, see Example: Understand a new repository (and the other walkthroughs such as Example: Build a new feature end to end) and the feature matrix.

Understand a codebase

cd the-repo
claude

> /init
> give me a high-level tour of this codebase: entry points, data flow, and the riskiest modules
> how does the billing flow work end to end?
> /compact keep the architecture summary

Tips:

/init writes a starter CLAUDE.md so future sessions start smarter (see Claude Code — Project Instructions (CLAUDE.md)).
Large explorations are delegated to the read-only Explore subagent automatically, keeping your main context clean (see Claude Code — Subagents).
Working across repos? /add-dir ../other-repo or claude --add-dir.
Use /btw for quick asides you don’t want in the conversation history.

Build a feature

claude --permission-mode plan

> /plan add pagination to the /users endpoint
# review the plan, approve it
> implement the plan, then run the tests
> /diff
> commit this with a conventional-commit message

Tips:

Plan mode is read-only until you approve the plan — the cheapest place to catch a wrong approach (see Claude Code — Permissions).
Shift+Tab toggles into plan mode at any time.
/diff shows uncommitted changes before you commit.
Long-running implementation? /background detaches the session as a background agent; check on it with /tasks.

Debug

cat failing-test-output.log | claude -p "find the root cause"

or interactively:

> the login flow returns a 401 after the token refresh — here's the failing test: [paste]
> form a hypothesis, add logging to confirm it, then fix it
> /rewind          # if the fix made things worse — rolls code and conversation back to a checkpoint

Tips:

/rewind (or Esc+Esc) rolls code and conversation back to a checkpoint — the fastest undo.
/debug [description] diagnoses issues with Claude Code itself; /doctor checks the installation.
Piped input (cat file | claude -p "...") is great for large logs that would flood the interactive UI.

Refactor

claude -w refactor-auth        # start in an isolated git worktree

> /plan extract the auth logic from middleware.ts into lib/auth/ without changing behavior
> implement it and keep the test suite green after each step

Tips:

--worktree (-w) isolates the refactor at <repo>/.claude/worktrees/<name> so your checkout stays clean.
For sweeping, repetitive changes, /batch <instruction> decomposes the work into units run in worktrees.
/branch forks an earlier conversation if you want to try two refactor strategies.

Code review

> /review 123                  # fast single-pass read-only review of PR 123
> /code-review high            # multi-agent review of the local diff
> /code-review high --fix      # same, but apply the suggested fixes
> /security-review             # vulnerability pass on the diff
> /pr-comments 123             # pull in reviewer comments to address

Tips:

/review, /code-review, and /security-review are distinct: quick read-only PR check, deep multi-agent diff review (low→ultra; ultra runs in the cloud), and security-focused pass.
Resume the session behind a PR with /resume (paste the PR URL) or claude --from-pr 123.

Automate with headless mode

# one-shot, machine-readable
claude -p "list the TODO comments in src/" --output-format json

# bounded and budgeted
claude -p "fix the failing tests" --max-turns 10 --max-budget-usd 2 --allowedTools "Bash(pnpm test *)" "Edit" "Read"

# chain runs via session IDs
claude -p "draft the migration" --output-format json    # grab session_id from the JSON
claude --resume <session_id> -p "now apply the review feedback"

Tips:

Built-in and user slash commands are not available in -p mode.
--json-schema gives validated structured output in the structured_output field.
See Claude Code — CLI Flags for the full automation flag set.

Last verified: 2026-09-07

Official sources:

AI coding tools evolve quickly. If you find outdated information, please open an issue or pull request.