Handoff To Claude Code

Handoff to Claude Code

Delegates work to a local Claude Code session through claude -p. The user's Claude subscription pays for it, so only run it when the user asked you to.

SKILL_DIR below is the directory holding this SKILL.md; the wrapper is $SKILL_DIR/scripts/handoff.py. Run python3 $SKILL_DIR/scripts/handoff.py doctor once on a new machine to confirm the CLI is installed and logged in.

Pick a mode

| | chat | agent | | --- | --- | --- | | The deliverable is | the text Claude writes | the work Claude does on disk | | Claude may | read, search, run commands | everything, including edits | | You should | relay the text verbatim | summarize what happened |

chat withholds Edit and Write, but it can still run shell commands - so it is the right mode when nothing should change, not a guarantee that nothing can.

python3 "$SKILL_DIR/scripts/handoff.py" chat  "Why does this crash on startup?"
python3 "$SKILL_DIR/scripts/handoff.py" agent "Port the auth module to the new API and run the tests"

Both commands print Claude's output, then a footer:

--- handoff metadata ---
session: 6b1c…   model: claude-opus-5   cost_usd: 0.12   duration_s: 41   dir: /path

The footer is for you, not the user. Never show it to them - except model:, which is the name to use in the attribution line below. It is the model that did the work, already disambiguated from the auxiliary model that permission checking runs in the background; take it as given rather than reasoning about token counts yourself.

For a long or multi-line prompt, pass - and pipe it in:

python3 "$SKILL_DIR/scripts/handoff.py" agent - 
python3 "$SKILL_DIR/scripts/handoff.py" tail    -n 40
python3 "$SKILL_DIR/scripts/handoff.py" wait    --timeout 600
python3 "$SKILL_DIR/scripts/handoff.py" kill   

kill stops the run and everything it spawned. Use it if the user changes their mind, or if a job is stuck - a job left running keeps spending the user's subscription. status reports died for a job that vanished without finishing.

Other options

• --model - override the model. Without it Claude Code uses the user's configured

default, which is usually what they want.

• --yolo - run with bypassPermissions instead of the default auto mode. Only when the user

asks for it or an ordinary run was blocked by permissions.

• --dir - run somewhere other than the current directory.
• --no-auth-check - by default the wrapper strips ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN

from the environment so the run bills the user's subscription rather than their API key. Pass this flag only if the user explicitly wants API-key or gateway billing.

• -- - anything after -- (and before the prompt) goes straight to claude.

When something fails

The wrapper prints [handoff] ERROR: … for failures that are not worth retrying - an expired login, an exhausted usage limit, a billing block. Report those to the user instead of retrying; they need a human to fix them. Other failures surface Claude's own stderr and exit status.

See reference.md in this directory for the full flag table, the permission model, and troubleshooting.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: milanglacier
• Source: milanglacier/handoff-to-claude-code-skill
• License: MIT

Install and usage instructions live in the source repository linked above.