Capafy logo Capafy
Get started
Autowin : One shot any task

Autowin : One shot any task

by: Raphael Vilain
We built Autowin by putting models in a loop: discover upgrades → test them → judge them with an assembly of adversarial sub-agents — and we ran that very process on itself, iterating the kit until each piece held up under its own review. The result: from one loosely-defined need, you get back work that has been built, tested, and scored by a panel of judge sub-agents — with closure authority kept outside the model (deterministic checks + you), so an "it's done" is never just taken on faith.
#Produtividade
Avaliação
São necessárias mais avaliações
Vendas
0
Como usar
Baixar

Autowin — a self-correcting work system for Claude Code (Windows).
A pipeline of skills + deterministic guardrails that move closure authority outside the model:
an "it's done" is never taken on faith — it is replayed and judged.

⚠️ Windows-only core. The deterministic enforcement layer (the PowerShell guardrail hooks) requires
Windows + Windows PowerShell 5.1+. On macOS/Linux the skills and constitution still load, but the hooks
do not fire — closure falls back to a self-declared state. Buy only if you run Windows.

The idea

Most agents drift: they certify "done" on text alone, chase off-topic polish, and rewrite their own rules on a bad diagnosis. Autowin puts the critical decisions in deterministic code (hooks) and with you — never in the model's self-judgment.

We built Autowin by putting models in a loop — discover upgrades → test them → judge them with an assembly of adversarial sub-agents — and we ran that very process on itself, iterating the kit until each piece held up under its own review.

The result: from one loosely-defined need, you get back work that has been built, tested, and scored by a panel of judge sub-agents.

Across ~50 teams, Autowin has driven reported productivity gains of 5–10× on work spanning development, design, and video / game creation (range reported by early teams — not a guarantee).

What you get

One installer skill (autowin-setup) that deploys the complete kit:

Component What it does
scout Surfaces a ranked shortlist of concrete improvement candidates on a target.
frame Defines the real need (the WHAT), surfacing the unstated; then explores the approach options.
terrain Prepares the autonomous self-correction ground (observability, environment, resume).
build Resolves one defect through to a green verified out-of-model (test / exit code / screenshot / query).
judge Adversarial, multi-lens external review, scored per dimension and looped to a quality threshold.
kaizen Turns a session's failures into human-approved improvements to the kit itself.
_engine The shared canonical mechanics every skill builds on.
17 PowerShell hooks + constitution The deterministic enforcement layer (stop-gate, fix-gate, git-auth-gate…) wired into Claude Code.

Chain: scout → frame → terrain → build → judge (build for a defect, kaizen for a recurring failure).

Requirements

  • Claude Code (CLI / desktop / IDE)
  • Windows + Windows PowerShell 5.1+ — required for the guardrail layer (the hooks are .ps1).
  • macOS/Linux: the skills + constitution load, but the hooks (enforcement) do not fire — you get the behavioral pipeline only, with self-declared closure.

Install

  1. Purchase & download — the package lands in your Claude Code skills directory (~/.claude/skills/autowin-setup).
  2. Run the installer (Windows PowerShell, from the autowin-setup folder):
    powershell -ExecutionPolicy Bypass -File .\install.ps1
    It unpacks the 6 skills + engine, installs the hooks, sets the constitution — with backups, idempotent re-runs, and a bundled uninstall.ps1.
  3. Restart Claude Code. Verify: & "$env:USERPROFILE/.claude/hooks/test-hooks.ps1".

📖 Full walkthrough in GETTING-STARTED.md. 🔒 Read SECURITY.md first.

How a run feels

  • You: "improve the onboarding flow" (vague).
  • Autowin: frames the real need → builds increments, each verified red→green out-of-model → a panel of judge sub-agents scores the result and sends defects back to fix → you get work that survived its own adversarial review, with every "green" backed by a real artifact.

Security

The hooks execute commands read from RUN.md files at closure time to replay a "green" state. Use Autowin only in trusted projects — a cloned RUN.md is untrusted input. Full-autonomy is opt-in (AUTOWIN_AUTONOMY=1) and off by default. Details in SECURITY.md.

FAQ

Does it work on macOS/Linux? The skills + constitution load, but the hook enforcement layer is Windows/PowerShell only — on non-Windows, closure is self-declared, not enforced.

Will it overwrite my config? It replaces ~/.claude/CLAUDE.md (original backed up to *.autowin-orig) and merges its hooks into settings.json, preserving your other keys and hooks. Same-named skills are backed up too. Reversible via uninstall.ps1.

Is the "5–10×" guaranteed? No — it's the range reported by early teams on their own work. Your mileage depends on task and setup.


View more