Scene 01 · Take 1 · Rolling

Claude doesn't doits own stunts.

Claude Code plans the task and reviews the result. A near-free model — DeepSeek, Groq, or local Ollama — does the typing through a local proxy. Your expensive subscription tokens are spent on thinking, never on typing.

scroll
The shot list

How the stunt is pulled off

Claude directs. The stunt double performs. Same footage, a fraction of the budget.

01

Plan

Claude reads your codebase, designs the change, and writes a precise spec — files, edits, acceptance checks.

Claude · subscription
02

Execute

The cheap worker model picks up the spec and writes every line of code, routed through a local proxy.

DeepSeek / Groq / Ollama
03

Review

Claude inspects the diff like a director at the monitor — catching bugs, style breaks, and missed cases.

Claude · subscription
04

Iterate

Notes go back to the worker for another take. Loop until the shot is clean — no extra thinking budget burned.

DeepSeek / Groq / Ollama
Dailies · real run

Footage from the cutting room

0
files shipped
a complete feature
0s
wall-clock time
plan to clean diff
0
review issues
passed first cut
$0
Anthropic tokens
on implementation
// not a benchmark — an actual stuntman run, captured end to end.
Overtime · second unit

Rolls right through your rate limit

Hit the 5-hour wall mid-scene? The stunt double keeps shooting on its own dime — and Claude picks up the moment the window reopens. Hands-free.

01

Detect

stuntman reads your 5-hour window — how much is spent, the exact reset — off the same meter /usage uses. Zero tokens, no API call.

window probe
02

Keep rolling

The stunt double isn't on Anthropic's clock. While the limit cools down, it keeps executing the spec'd work in the background — your key, not your cap.

DeepSeek / Groq / Ollama
03

Resume

The window reopens and Claude wakes where it left off — reviews the night's footage, rolls the next take. Auto when the reset is close; one ping when it's not.

Claude · subscription
Continuity · next call

Picks up on the next take

Hit the context limit, or just start fresh tomorrow? The project keeps its own notes — the next session reads them first and rolls, with zero re-explaining.

01

Scaffold once

One /scaffold writes a contract into your CLAUDE.md and creates four living docs — HANDOFF, STATUS, SPEC, STRATEGY — idempotent, never clobbers.

one-time setup
02

Wrap the take

Before each session ends, the agent refreshes HANDOFF.md, STATUS.md, and the README — what changed, what's next, the gotchas. The project becomes its own memory.

every session
03

Cold open

New session, cleared context, tomorrow morning — /handoff reads the docs first. You just say "execute handoff" and it rolls.

zero re-explaining
The call sheet · every command

Six commands, one crew

Plan → execute → review is the core. The rest make long runs cheap, rate-limit-proof, context-proof, searchable across every project — and turn a blank page into a launch plan.

/delegate
Plan with Claude, execute with a near-free model, review with Claude. Your subscription tokens go only to thinking. — saves cost
/relay
Span Claude's 5-hour usage limit. The stunt double keeps executing while you're capped; Claude resumes at reset. — beats the rate limit
/scaffold
Give a project self-resuming memory. A CLAUDE.md contract + four living docs (HANDOFF · STATUS · SPEC · STRATEGY). — survives the context boundary
/handoff
Resume exactly where the last session stopped. Reads the living docs first — zero re-explaining, even after /clear. — cold-opens any session
/wiki
Build a "second brain" across a folder of projects. An Obsidian vault of notes-per-project + a graphify knowledge graph + a live MCP — so any session can ask "did I solve this already?" — cross-project memory · new in v0.7
/launch
Turn a blank page into a product launch plan. A multi-agent workflow does cited competitor research, sizes the market, ranks channels, then writes pricing, positioning + a week-by-week playbook — pressure-tested by adversarial critics — into one Product Success Overview (markdown + HTML). — go-to-market · new in v0.8
Action

Two takes and you're rolling

It's a Claude Code plugin. Open source. No keys to swap, no config to babysit.

claude-code — stuntman
/plugin marketplace add mhlaghari/stuntman
/plugin install stuntman@stuntman

Then just work as usual — Claude calls in the stunt double whenever there's typing to be done.