Claude Code Skill Guide

Workflow

1. Confirm task mode:

- New run: use claude -p. - Continue prior run: use claude -c -p with stdin prompt. - Resume specific session: use claude -r <session-id> -p.

1. Choose session strategy before picking other flags:

- Ephemeral one-shot run: include --no-session-persistence (default). - Resumable run: omit --no-session-persistence only when the user is likely to continue or resume later.

1. Set defaults unless user overrides:

- Model: sonnet (default). - Effort: high (default for delegated work). - Permission mode: plan unless edits are required.

1. Build command with required flags:

- Always include -p (print mode) for non-interactive execution. - Include exactly one session flag path: plain -p, -c -p, or -r <session-id> -p. - Include --no-session-persistence only for ephemeral runs.

1. For long or multi-line prompts, prefer stdin over argv so large prompts do not hit shell-escaping or argument-length issues:

- cat > /tmp/claudeprompt.txt <<'EOF' ... EOF - claude -p ... < /tmp/claudeprompt.txt

1. Run command and validate output quality:

- If exit code is non-zero: handle as failure. - If exit code is zero but stdout is empty/whitespace: treat as a soft failure and retry once using the stdin prompt pattern plus an explicit output contract (for example: "Reply in sections 1..N").

1. Summarize outcome and ask what to do next.
2. After completion, if session persistence was enabled for the run, remind user they can continue with claude -c -p or resume a specific session with claude -r <session-id> -p.

Quick Reference

Command Patterns

New run

claude -p \
  --model sonnet \
  --effort high \
  --permission-mode plan \
  --no-session-persistence \
  "your prompt here"

New run (robust for long prompts)

cat > /tmp/claude_prompt.txt <<'EOF'
your long multi-line prompt here
EOF

claude -p \
  --model sonnet \
  --effort high \
  --permission-mode plan \
  --no-session-persistence \
  < /tmp/claude_prompt.txt

New run with edits

claude -p \
  --model sonnet \
  --effort high \
  --permission-mode acceptEdits \
  --no-session-persistence \
  "your prompt here"

Continue most recent session

echo "your follow-up prompt" | claude -c -p

When continuing, do not add configuration flags unless the user explicitly asks for changes (for example, different model or effort level).

Resume specific session

echo "your follow-up prompt" | claude -r <session-id> -p

Run in a worktree

claude -p -w \
  --model sonnet \
  --effort high \
  --permission-mode acceptEdits \
  --no-session-persistence \
  "your prompt here"

Budget-capped run

claude -p \
  --max-budget-usd 5.00 \
  --model sonnet \
  --permission-mode acceptEdits \
  --no-session-persistence \
  "your prompt here"

Tool-restricted run

claude -p \
  --allowedTools "Read Grep Glob Bash" \
  --permission-mode plan \
  --no-session-persistence \
  "your prompt here"

Structured output

claude -p \
  --output-format json \
  --json-schema '{"type":"object","properties":{"summary":{"type":"string"},"issues":{"type":"array","items":{"type":"string"}}},"required":["summary","issues"]}' \
  --no-session-persistence \
  "your prompt here"

Model Options

sonnet is the default for software engineering tasks.

Session Strategy

• Default to ephemeral one-shot runs with --no-session-persistence for delegated tasks, scripts, and automations.
• Use persistent sessions only when the user explicitly wants follow-up turns, resume support, or iterative collaboration within Claude itself.
• Do not combine "resumable" guidance with ephemeral flags in the same example. Pick one session strategy and keep the command shape consistent.

Effort Levels

• high - Default for delegated work (deep analysis, thorough implementation)
• medium - Standard tasks (feature additions, bug fixes)
• low - Simple tasks (quick fixes, formatting, documentation)

Permission Modes

Treat the Codex sandbox column as a rough mental model, not a literal one-to-one mapping. The CLIs expose different controls and approval semantics.

Following Up

• After every run, ask for next steps or clarifications.
• When proposing another run, restate model, effort level, and permission mode.
• For continuation, use -c -p with stdin when the earlier run used session persistence.

Error Handling

• If claude --version or claude -p exits non-zero, report failure and ask before retrying.
• Ask permission before high-impact flags unless already granted: --permission-mode dontAsk, --dangerously-skip-permissions.
• If output includes warnings or partial results, summarize and ask how to proceed.
• If claude -p exits 0 but returns empty output, rerun once with the robust stdin prompt pattern and a stricter output instruction; if still empty, report the run as failed/indeterminate.

CLI Version

Use a current Claude Code version. Check with:

claude --version