Claude Code

Overview

Run Claude Code from the terminal to analyze, plan, refactor, or edit codebases via the

claude

CLI. Default to non-interactive print mode in this environment and summarize results back to the user.

Quick Start

1. Use non-interactive (print) mode:

   claude -p "prompt"

   for one-shot tasks. Interactive sessions are disabled here due to flaky TTY prompts.
2. Select model (ask the user). Default suggestion:

   opus

   .
3. Choose permissions and scope.
4. Run the command, capture output, and summarize key results and next steps.

Workflow

1) Ask for model and session intent

• Ask the user which model to use.
• If the user doesn't specify, propose

  --model opus

  and confirm.

• --model

  accepts

  sonnet

  or

  opus

  aliases, or a full model name.

2) Pick the right command form

Non-interactive / print mode (default):

• claude -p "prompt"

  to run a single request and exit.

• cat file | claude -p "prompt"

  to pipe content in.
• Use

  --output-format json

  or

  --output-format stream-json

  for machine-readable output when scripting.
• Use

  --max-turns N

  to limit agentic turns in print mode.
• Use

  --input-format

  for

  text

  or

  stream-json

  inputs.
• Use

  --no-session-persistence

  to avoid saving sessions in print mode.
• Use

  --include-partial-messages

  with

  --output-format stream-json

  to emit partial streaming events.

Session management (print mode only):

• Continue the most recent session in print mode:

  claude -c -p "prompt"

  .
• Resume a specific session in print mode:

  claude -r "<session-id>" -p "prompt"

  .

• --fork-session

  creates a new session ID when resuming.

3) Permissions and safety

• Prefer default permission prompts. Only use

  --dangerously-skip-permissions

  if the user explicitly approves and the task is low-risk or sandboxed.
• Use

  --permission-mode

  to start in a specific mode and

  --allow-dangerously-skip-permissions

  only to permit the bypass option.
• For larger changes, explore first, then plan, then implement.
• Use

  --allowedTools

  /

  --disallowedTools

  or

  --tools

  to control tool access when needed.

4) Run, summarize, and follow up

• Execute the command, capture stdout (and stderr if needed).
• Summarize results, list changes or findings, and ask the user for next steps.

Notes

Assume the user already has Claude Code installed and configured. If

claude

is not found, report the error and ask the user to fix their environment before retrying. When prompting, prefer rich context: reference files with

@

and pipe logs.

Troubleshooting

If the prompt passed to

--print

/

-p

is misinterpreted (e.g., flags inside the prompt), use

--

before the prompt or pipe it via stdin:

• claude -p -- "prompt with -dashes or --flags"

• echo "prompt" | claude -p