Offload
Runs a target skill in a remote cloud agent and writes the output to the local workspace.
Inputs
/offload <target-skill> [input-content]
<target-skill>— the name of the skill to invoke remotely (e.g.to-seed,to-tasks).[input-content]— optional content to pass as input to the remote skill. If omitted, compress the current conversation context into a brief (a few hundred words) and use that as input.
Process
1. Parse args
Read the user's message. Extract target-skill and any input-content exactly as provided.
If no target skill is given, print the invocation form above and stop.
If no input-content was provided, compress the current conversation to a brief:
- Include the topic, key decisions, and any file paths or data needed by the target skill.
- Keep it under 400 words — enough context for a cold agent to run the skill without this session.
- Always include the current git branch: run
git branch --show-currentand appendCurrent branch: <result>to the brief. Skills liketo-seeduse this to skip their interactive branch prompt.
2. Look up model and effort, and read the target SKILL.md (parallel)
Issue both reads in the same turn — they are independent:
- Read
~/.dotfiles/claude-code-shared/resources/model-tiers.jsonto resolve the tier fortarget-skill(fall back todefaultif not listed). Resolvemodelandeffortfrom thetierssection. - Read
~/.dotfiles/claude-code-shared/skills/<target-skill>/SKILL.md. If the file is not found, stop and report the missing path. Do not guess alternate locations.
Wait for both reads before step 4.
4. Spawn the remote agent
Spawn an Agent with:
isolation: "remote"
model: <resolved model>
description: "Offloaded <target-skill> run"
prompt: <see template below>
Prompt template:
You are running the /<target-skill> skill at effort level: <resolved effort>.
## Skill instructions
Execute the following skill exactly. Do not deviate from its output format.
<full contents of the target SKILL.md, pasted verbatim>
## Input
<input-content or compressed context brief>
## Remote execution constraints
You are running cold with no access to the local filesystem or session transcript. Apply these rules:
1. **Branch prompts**: do not ask. Use the `Current branch:` value from the Input section above.
2. **Adversarial verification** (e.g. to-seed persona stage): skip it. Set `verification.status: "degraded"` and `verification.skipped_reason: "no local transcript available in remote execution"`. Do not spawn persona sub-agents.
3. **File writes**: do not write files. Return the output as text — the caller writes it to disk.
## Output instructions
1. Follow the skill's output format exactly — same structure, same labels, same ordering.
2. Return ONLY the skill's final output in your response — no truncation, no summary.
- If the skill outputs plain text (e.g. pr-code-review), return plain text.
- If the skill outputs a JSON file (e.g. to-seed, to-tasks), return valid JSON.
- Do NOT invent a wrapper format. Return verbatim what the skill specifies.
Effort level reminder: work at <resolved effort> effort — <effort description>.
Effort descriptions by level:
low— fast lookup, minimal reasoning, answer directly.medium— mechanical execution, well-defined scope, no open-ended exploration.high— context-aware, standard depth.xhigh— deep reasoning, architecture-level thinking.
5. Receive output and write to disk
Receive the agent's text response.
Determine the output directory and file extension:
- If the target skill produces seeds (e.g.
to-seed): write todocs/seeds/, usejson - If the target skill produces tasks (e.g.
to-tasks): write todocs/tasks/, usejson - Otherwise: write to
docs/offload-output/(create if needed), usemd
Generate a filename:
~/.dotfiles/claude-code-shared/scripts/scaffolding/doc-filename.sh <target-skill>-offload <ext>
Write the agent's response verbatim to that path.
6. Display the output
After writing, display the agent's output directly to the user — exactly as the skill would have shown it locally. Do not summarize or reformat. Print the full content, then report the output path on a final line:
Output: <path>
Read and execute ~/.dotfiles/claude-code-shared/resources/learning-capture.md.
This skill's slug is offload.