AI 工程从零到一（中文）

Given a repo, an agent product, and a small sample app, produce a portable evaluation harness that compares prompt-only against workbench-guided pipelines.

Produce:

1. eval/sample_app/ — a minimum-viable sample app drawn from the project's domain.
2. eval/run_prompt_only.py and eval/run_workbench.py that each take a task description and return a TaskOutcome.
3. eval/report.py that runs both pipelines and writes before-after-report.md plus comparison.json.
4. CI workflow that fails when workbench outcomes regress on a fixed task suite.
5. docs/benchmark.md explaining the five outcomes and what counts as a regression.

Hard rejects:

• A benchmark with only one pipeline. Comparison is the whole point.
• Outcomes phrased as percentages without a denominator. Always report n / m.
• A sample app the agent product was trained on. Use a domain-tuned fixture.
• Reports that hide false negatives. Tasks where prompt-only was faster must be enumerated.

Refusal rules:

• If the project has no acceptance command, refuse to ship the benchmark. There is nothing to measure.
• If the workbench pipeline takes more than 3x the prompt-only pipeline on the median task, surface that finding; the workbench needs simplification, not the model.
• If the harness cannot run offline, refuse to wire it into CI. Network flakiness will corrupt the comparison.

Output structure:

<repo>/
├── eval/
│   ├── sample_app/
│   ├── run_prompt_only.py
│   ├── run_workbench.py
│   └── report.py
├── outputs/eval/
│   ├── before-after-report.md
│   └── comparison.json
├── docs/benchmark.md
└── .github/workflows/benchmark.yml

End with "what to read next" pointing to:

• Lesson 42 for the capstone pack that bundles every surface used by the workbench pipeline.
• Lesson 19 (SWE-bench, GAIA, AgentBench) for the macro benchmarks this complements.
• Lesson 30 (Eval-Driven Agent Development) for ongoing eval loops once the benchmark is wired.