The actual rules, skills, hooks, and scripts I hand to my AI agents — the working artifacts behind my build-in-public series My Wise Buddy and I.
I'm not a developer. I work a day job and build a one-person AI company in the evenings. Every article in the series describes something I actually shipped; this repo is where the shipped thing lives, so you can read it instead of taking my word for it.
Language note: I work in Korean, so the pack documents are Korean. Each pack gets an English summary here in the README. Code, file names, and CLI flags are language-neutral.
Two governance components for working with coding agents, plus the hook and scripts that measure whether they're working. Referenced by #2: Deciding Less, Twice.
1. autonomy-calibration — where the line between "ask me" and "just decide" should sit, and how to move it with data instead of vibes.
- Two-phase operation: a dense agreement phase up front (problem definition, success metrics, review protocol, irreversible boundaries) → an 8-item gate → the agent proceeds on its own.
- A one-page autonomy contract template that the gate is scored against.
- A question router that classifies each question the agent is about to ask into one of four paths (agent decides / verify from sources / ask the human / stop and report), plus a PreToolUse hook that reminds the agent right before it asks.
- Measured calibration: autonomous-reversal rate (how often the human overturns agent-only decisions → narrow autonomy) and unnecessary-question rate (how often it asked what it could have resolved → widen autonomy). Defaults: 10% / 30%.
- Protected questions that no optimization may skip: irreversible or outward-facing actions, unclear internal facts, the human's own preferences, unresolved trade-offs, security and legal matters.
2. project-three-notes — three notes per project, each with one authority.
- Fact note: what is true, with sources, verification status, and a re-check date.
- Decision note: why we chose this — including who decided (6 kinds, from "human proposed" to "agent decided alone") and a later verdict (held / overturned / discarded). The pair is the point: counting agent decisions tells you nothing; agent decisions × later overturned tells you where to narrow.
- Prompt note: generated from transcripts by script, indexed by session, secrets masked, full text opt-in and git-excluded by default.
How v2.0 was built: the same instructions were implemented twice — once by the session that had lived through the problem, once by a clean-room session that had never seen it. The context session won the core design; the clean room won instrumentation and enforcement. v2.0 is the merge, and appendix/COMPARISON.md is the scorecard. Both reviews (a fresh-context subagent and an independent model) are logged there too.
Install: autonomy-pack/INSTALL.md (Claude Code on Windows; PowerShell + Python 3.10+). Prefer the offline snapshot? autonomy-pack_v2.0.zip is the same tree, zipped.
Heads up: this pack points at a few rules from my personal governance set that aren't included here (decision-presentation, consultation-default, scope-discipline, verification, credentials-handling, progressive-disclosure). It works without them — those pointers just go quiet.
MIT — see LICENSE. Use it, fork it, strip it down. If it saves you a mistake I already made, that's the whole point.