Prime Worker — Development skill for Claude Code

prime-worker

A [Claude Code](https://claude.com/claude-code) skill for delegating **multi-turn** work to a local [`prime-agent`](https://github.com/PrimeIntellect-ai/prime-agent) subagent that remembers context across turns.

Claude spawns a cheap model as a named worker, keeps talking to it, branches it, and lets it learn from its own mistakes — while the expensive context stays free for the work that needs it.

pw spawn auth-review --cwd ~/proj/src/auth --model deepseek-v4-flash --cite "Map the token refresh path."
pw say auth-review "Now trace where the refresh token is persisted."
pw fork auth-review auth-alt "What breaks if the TTL drops to 60s?"
pw refine auth-review          # let it record what it learned
pw retire auth-review --yes

The idea: a worker is a named directory

`~/.claude/prime-workers//` holds one session `.jsonl` plus a `worker.env` recording the `--cwd`, `--model` and flags it was spawned with. The name is the only handle you need — no session ids, no `ls -t` races, and concurrent workers cannot collide.

That matters more than it sounds. Sessions are UUID-named `.jsonl` files, so without a convention you end up carrying ids between turns, racing `ls -t` whenever two workers run at once, and guessing which of thirty files in the store is the 50MB one. A named directory gives each worker a stable, readable handle, lets concurrent workers coexist, records the flags it was spawned with so follow-up turns need none, and makes retiring one an `rm -rf`.

Requirements

• `prime-agent` on your PATH (developed against 0.8.0)
• A configured model provider. Defaults target an OpenCode Go subscription with deepseek-v4-flash; override with PW_PROVIDER and PW_DEFAULT_MODEL for any provider prime-agent supports.
• bash, python3

Install

git clone https://github.com/alperiox/prime-worker ~/.claude/skills/prime-worker
~/.claude/skills/prime-worker/scri