05-core-delegate-to-cheaper-llm-via-shelllisted
Install: claude install-skill markkennethbadilla/public-agent-provisioning
# Delegate to a cheaper LLM — via SHELL (SANCTIONED SUBAGENT TOOL)
## The mandate (why this skill is HARD)
Native subagents (`Task`/`Agent`) are **banned** (rule 20c) — they don't share your
context and can't use a custom model/key. **A shell call to a cheap model IS your
subagent tool.** Any time you'd reach for a subagent — bounded reasoning, a second
opinion, "draft/summarize/classify/extract this," a focused unit of thinking — you
delegate it **here, from your main thread**. CLAUDE.md, the gates, and the vault stay
intact (it's just a model call; YOU still own the work, edits, deploys).
**Never underutilize it.** "I'll just reason through this myself when a cheap model
could carry it" is a missed delegation — the same defect as reading a whole file
instead of using jcodemunch.
### Why SHELL, not an MCP (this is the whole point)
The old `deepseek-delegate` MCP took the prompt as a string param, so the payload
**always passed through your context** and burned your tokens. A shell script lets the
payload be **piped in, unread by you**:
```
Get-Content big.txt | bun delegate.mjs "summarize the decisions" # PowerShell
cat big.txt | bun delegate.mjs "summarize the decisions" # bash
bun delegate.mjs "summarize the decisions" --file=big.txt # or --file
```
The bytes flow stdin/file → HTTP → model; **only the short reply returns to you.**
### What you DO vs DON'T delegate
- **DELEGATE:** bounded reasoning, drafting, summarizing, c