AI Engineering from Scratch（英文原版）

Given a repo and any existing prose instructions (AGENTS.md, CONTRIBUTING.md, onboarding docs), produce a five-category rule set the workbench can execute.

The five categories:

1. startup — what must be true before work begins.
2. forbidden — what must never happen.
3. definition_of_done — what proves the task is complete.
4. uncertainty — what the agent does when not sure.
5. approval — what requires human sign-off.

Produce:

1. docs/agent-rules.md with one ## heading per rule. Each rule carries category, check, and a one-line description.
2. tools/rule_checker.py with a RuleChecker class exposing one method per check. Each method takes a TurnTrace dataclass and returns bool.
3. tools/rule_report.py runner that loads rules, runs the checker on a trace, emits a rule_report.json.
4. A migration notes file: which prose lines became which rule, which were dropped as aspirational, why.

Hard rejects:

• Rules without a check field. Aspirational-only rules belong in onboarding docs, not in the workbench rule set.
• A single "be careful" rule. Specify a category and a check or remove it.
• Checks that require LLM calls. Rule checks must be deterministic and cheap so they can run every turn.
• Rule files over 200 lines. Split by category into agent-rules.\{startup,forbidden,done,uncertainty,approval\}.md and route from a parent index.

Refusal rules:

• If the agent product cannot supply a TurnTrace (no instrumentation), refuse to wire the checker until at least read_state_file, edited_files, and tests_exit_code are recorded.
• If existing instructions are mostly aspirational (>50%), surface that finding before emitting rules. The rule set will look thin; that is correct.
• If a rule is added because of a single past incident, attach the incident id so future review can decide if it is still needed.

Output structure:

<repo>/
├── docs/
│   └── agent-rules.md
├── tools/
│   ├── rule_checker.py
│   └── rule_report.py
└── docs/migration-notes.md

End with "what to read next" pointing to:

• Lesson 36 for per-task scope contracts that extend the forbidden category.
• Lesson 38 for verification gates that consume the rule report.
• Lesson 39 for the reviewer agent that scores rule compliance.