Curation Discipline

curation-discipline

**Keep an AI agent from silently corrupting your dataset.**

繁體中文版：[README.zh-TW.md](README.zh-TW.md)

A Claude Code plugin: five agents, three skills, two tested hooks, and a catalogue of data-pipeline failures that produce no error message.

---

The problem

You point an agent at a database and ask it to help you build a dataset. It works. The rows appear. Some months later you discover that the same entity exists twice under two names, a batch of rows has no attribution, and a sync has been publishing a stale mirror since a schema change nobody connected to it.

None of that produced an error. That is the actual problem — **not that agents make mistakes, but that the mistakes they make are the quiet kind.** A crash tells you where and when. A join that matched nothing tells you nothing at all, and by the time you notice, the bad rows are indistinguishable from the good ones.

What this is

One rule, and the machinery that keeps people from working around it:

**No agent writes to the canonical store. Agents write to staging. A human moves staging into the store.**

  collect            check              approve           write
  ────────           ─────              ───────           ─────
  curator     →     validator     →     human      →     human
  (writes           (read-only,         (reads the       (runs the
   staging)          hook-enforced)      diff)            SQL)

The staging file is the review artifact. It is the difference between "the agent added 40 rows" and "here are 40 rows, 3 need your attention, and here is why".

Install

/plugin marketplace add jerryliu0103/curation-discipline

/plugin install curation-discipline@curation-discipline

Traditional Chinese edition: `curation-discipline-zh@curation-discipline`. Same content, same structure — install one, not both.

What you get

Agents

Agent	Job	Can write
curator	Collects and normalizes