SystemWright

See it in 30 seconds — illustrative walkthrough

You give it — raw, messy

· 28 leads this week (+6), but only 3 reached a quote
· Customer B needs a delivery date confirmed by Friday
· "optimize the review flow" was decided — but nobody owns it
· support hit the same question 5× this week
"help me run my company with AI"

It returns — a runnable loop, not a platform

facts ≠ opinions · 2 concrete decisions · human-approval gates · next-week checks
→ Decision 1: assign an owner for the review flow today; minimal plan by Thursday.
→ Decision 2: give Customer B a firm date by tomorrow noon — don't let it slip.
→ Action: send 2 messages in the next 10 minutes; write 5 facts, not a report.

No platform built. No tool forced. Just the smallest durable system that moves the real goal forward — with you still deciding what matters. Illustrative example — the full recorded design walkthroughs (honestly labelled self-authored) live in the repo; the saas-support design has additionally been replayed over 100 real labeled tickets (raw data in the repo).

Four layers, not one prompt

Prompt

Say it clearly

Role, task, constraints, output format, and the decisions the AI may make alone vs. must escalate.

Context

Give the right facts

The minimal high-signal set: files, examples, rules, memory, and a single source of truth.

Harness

Bound the action

Tools, permissions, sandbox, verification commands, and what stays read-only.

Loop

Make it improve

Trigger, review cadence, memory, stop conditions — and the forbidden means that stop it gaming its own check.

Install in one line

It's a plain SKILL.md directory — installation is copying it into your host's skills folder. The bundled installer does it portably:

# pick your host: claude | codex | gemini | cursor
git clone https://github.com/EthanQC/SystemWright && cd SystemWright
sh scripts/install.sh --host claude

Then just describe a recurring goal in your agent. Full per-host table and Kimi/GLM setup in the README.